GAIA: Portuguese-language AI Built for Brazil
GAIA is part of a growing wave of localized AI trained on Latin American languages, use cases, and institutions.
Latin America’s push to create artificial intelligence that reflects its own languages and culture took a major step forward with the release of GAIA, a large language model trained entirely in Portuguese.
GAIA, which appears to be a reference to a 14th-century Portuguese legend, was developed by Google’s AI research lab DeepMind in collaboration with a Brazilian university and two startups.
It is built on Gemma 3, Google’s open-source AI model, optimized for Brazilian Portuguese and designed for local use cases such as service chatbots.
DeepMind in a statement said that GAIA is already being piloted by the Tribunal de Contas dos Municipios de Goiás (TCM-GO), an audit institution that monitors the finances of municipal governments in the state of Goiás.
In initial trials of the LLM, the TCM-GO’s technology manager described it as providing better context than other models when answering questions while offering fewer extraneous details.
Health insurer Unimed is exploring using GAIA to validate diagnoses of spine injuries, while financial advisory firm BHub is looking at using it for customer service.
University and startup partners
A key partner in the project is the Center of Excellence in Artificial Intelligence (CEIA) at Universidade Federal de Goiás. The state of Goiás has been aggressively seeking to become a center of Brazil’s AI industry, and last month became the first Brazilian state to create its own artificial intelligence regulations.
Two Brazilian startups contributed its development - Nama, which creates proprietary chatbots for the state of Sao Paulo, Canon Brazil, and retailer Magazine Luiza; and Amadeus, which is focused on Portuguese-language AI solutions.
Separately, Instituto de Ciência e Tecnologia Itaú-Unibanco (ICTi) - the research center for Brazil’s largest bank - released an open-source Portuguese-language dataset called Aroeira designed to support AI training.
Too soon for a test drive
Sadly I was unable to try it out. GAIA is currently available only as a model rather than a full-service AI application, which is akin to having an engine without a car.
AI enthusiasts with more computing power than my modest laptop, or a paid plan on Hugging Face - a machine learning that lets users build, deploy and train machine learning models - might have more luck. (I was similarly unable to make it work on Google’s Vertex AI platform, where it’s also available.)
What kinds of questions would a Brazilian LLM answer better than a mainstream chatbot? I’d want its recommendations for someone (Brazilian or foreigner) moving to Rio de Janeiro to help them navigate the city’s famously cliquey social scene. Or how to manage traffic on Sao Paulo’s Marginal Pinheiros? Or maybe a detailed vacation plan for traveling from Jericoacoara to the Lençóis Maranhenses by jeep?
What would you ask GAIA?
Why local models matter
Mainstream LLMs like ChatGPT, Gemini, and Claude are trained mostly in English. That means when Latin Americans use them in Portuguese or Spanish, it’s a bit like watching a dubbed Hollywood movie: the words translate, but the nuance gets lost.
As Mexican AI expert León Palafox put it: “The artificial intelligence that is revolutionizing the world thinks in English, and barely babbles our language.”
Around 90% of GPT-3’s training materials were in English, with less than 1% in Spanish, he notes, adding that chatbot performance on benchmark evaluations drops significantly when those evaluations are translated into Spanish.
These chatbots don’t understand Latin American cultural context, meaning they could easily confuse the significance of words like “güevón” in Venezuela or “güey” in Mexico which in a friendly context are similar to “dude” but in an adversarial one could easily start a fight.
Latin Languages
Chile is advancing the development of a Spanish-language LLM, with Chile’s National Center for Artificial Intelligence, or CENIA, expecting to release a model to the public in September.
“If we want to be part of the debate on governance, adoption, and putting AI at the service of people in our countries … we have to get our hands dirty,” CENIA director Rodrigo Durán told Foreign Policy’s Catherine Osborn last week.
Models like Gaia and LatAm-GPT won’t rival ChatGPT in raw computing power or programming fluency - and they don’t need to. Latin American engineers and coders who use AI for complex calculations or software development can stick with mainstream LLMs.
But for use cases like customer service, legal document drafting, or public-facing AI tools, these localized models offer crucial advantages.
LatAm AI Headlines
Mexican film producers have formed an alliance to address AI's impact on cinema in the hopes of protecting jobs and preserving cultural identity in filmmaking, reports Vanguardia.
Brazilian AI influencer Marisa Maiô, who has effectively become an artificial-intelligence-powered talk-show host, has gained enormous social media popularity and has now signed a partnership with retailer Magazine Luiza, reports Exame.
For more on AI in Brazil, listen to Anderson Soares and Fabio Cozman - two of Brazil’s most respected AI experts - on Estadão’s Dois Pontos podcast.
Macro Prompt
Bigger. Not smarter.
Apple last week released a paper that calls into question the excited predictions that AI is on the cusp of developing human-level intelligence.
LLMs provide statistically derived combinations of words through mathematical models trained on huge amounts of text. The staggering leaps in what AI can do in recent years are the result of training them on bigger and bigger datasets - but this isn’t the same as teaching them to reason through a problem.
Large Reasoning Models, or LRMs, are the next step in an effort to move beyond “stochastic parrots” into machines that think and understand.
This isn’t going well, according to an Apple Machine Learning Research paper called The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity.
“Frontier LRMs face a complete accuracy collapse beyond certain complexities. Moreover, they exhibit a counter-intuitive scaling limit: their reasoning effort increases with problem complexity up to a point, then declines,” reads the paper.
This contrasts with the euphoric - or perhaps dystopian - predictions that we are within striking distance of Artificial General Intelligence - or AGI - AI that can learn, reason, and adapt across any task like a human.
I personally find AGI a somewhat fuzzy concept typically accompanied by AI-related hyperventilation of both utopian and doomsday extraction. Yes, I believe AI will change the world. How? And how fast? Right now, I don’t think we have any way of knowing.