Latam-GPT is new massive language mannequin being developed in and for Latin America. The undertaking, led by the nonprofit Chilean Nationwide Heart for Synthetic Intelligence (CENIA), goals to assist the area obtain technological independence by creating an open supply AI mannequin educated on Latin American languages and contexts.
“This work can’t be undertaken by only one group or one nation in Latin America: It’s a problem that requires everybody’s participation,” says Álvaro Soto, director of CENIA, in an interview with WIRED en Español. “Latam-GPT is a undertaking that seeks to create an open, free, and, above all, collaborative AI mannequin. We’ve been working for 2 years with a really bottom-up course of, bringing collectively residents from completely different international locations who wish to collaborate. Just lately, it has additionally seen some extra top-down initiatives, with governments taking an curiosity and starting to take part within the undertaking.”
The undertaking stands out for its collaborative spirit. “We’re not seeking to compete with OpenAI, DeepSeek, or Google. We wish a mannequin particular to Latin America and the Caribbean, conscious of the cultural necessities and challenges that this entails, similar to understanding completely different dialects, the area’s historical past, and distinctive cultural elements,” explains Soto.
Because of 33 strategic partnerships with establishments in Latin America and the Caribbean, the undertaking has gathered a corpus of information exceeding eight terabytes of textual content, the equal of tens of millions of books. This data base has enabled the event of a language mannequin with 50 billion parameters, a scale that makes it similar to GPT-3.5 and offers it a medium to excessive capability to carry out advanced duties similar to reasoning, translation, and associations.
Latam-GPT is being educated on a regional database that compiles data from 20 Latin American international locations and Spain, with a formidable whole of two,645,500 paperwork. The distribution of information reveals a major focus within the largest international locations within the area, with Brazil the chief with 685,000 paperwork, adopted by Mexico with 385,000, Spain with 325,000, Colombia with 220,000, and Argentina with 210,000 paperwork. The numbers mirror the scale of those markets, their digital growth, and the supply of structured content material.
“Initially, we’ll launch a language mannequin. We count on its efficiency typically duties to be near that of huge industrial fashions, however with superior efficiency in subjects particular to Latin America. The thought is that, if we ask it about subjects related to our area, its data might be a lot deeper,” Soto explains.
The primary mannequin is the place to begin for creating a household of extra superior applied sciences sooner or later, together with ones with picture and video, and for scaling as much as bigger fashions. “As that is an open undertaking, we wish different establishments to have the ability to use it. A gaggle in Colombia may adapt it for the college schooling system or one in Brazil may adapt it for the well being sector. The thought is to open the door for various organizations to generate particular fashions for explicit areas like agriculture, tradition, and others,” explains the CENIA director.
