Latam-GPT is an innovative large language model under development to meet the specific needs of Latin America. Spearheaded by the Chilean National Center for Artificial Intelligence (CENIA), its goal is to foster technological independence by creating an open source AI model that reflects the diverse languages and cultural contexts of the region.
Álvaro Soto, the director of CENIA, emphasizes the collaborative nature of the project, stating that achieving this goal requires participation from multiple countries and communities across Latin America. Latam-GPT distinguishes itself from leading models created by multinational tech firms by focusing on the unique linguistic and historical elements of Latin America and the Caribbean. The initiative has been operational for two years and has recently gained traction with increased governmental involvement.
In forging ahead with the project, CENIA has formed 33 strategic partnerships across Latin America and the Caribbean, assembling a substantial database exceeding eight terabytes of text—equivalent to millions of books. The resulting language model consists of 50 billion parameters, placing it on par with GPT-3.5 and enabling it to effectively tackle complex tasks such as translation and reasoning.
Latam-GPT is being fine-tuned on a dataset compiled from 20 countries, accumulating a total of 2,645,500 documents. The distribution of this data highlights the representation from larger nations, with Brazil contributing the most at 685,000 documents, followed by Mexico, Spain, Colombia, and Argentina. The focus is on ensuring regional relevance and enhancing knowledge in subjects pertinent to Latin America.
The first model rollout is slated for this year and aims to deliver performance comparable to commercial models, especially in Latin American contexts. According to Soto, its ultimate purpose is to serve as an adaptable tool, allowing entities within the region to tailor the model to distinct sectors like education and health.
The initiative is bolstered by supercomputing resources at the University of Tarapacá in Chile, which has invested $10 million in infrastructure to support the model’s development. This groundbreaking capacity will enable local large-scale training for the first time, promoting technological decentralization.
As Latam-GPT evolves, it is set to expand horizontally to incorporate video and image functionalities, with future aspirations to integrate indigenous languages through collaborative efforts to develop translators.
Soto highlights the significance of developing a model grounded in local realities, particularly in education, asserting that it is critical for young people in Latin America to have tools that reflect their heritage and cultural narratives. In conclusion, for Latam-GPT to be deemed successful by 2030, it must facilitate the creation of localized AI applications that not only cater to but also empower the region’s diverse populations.
For more information, you can visit Wired en Español.
