Inside LatAm-GPT: Chile leads regional effort to build culturally relevant AI with a Latin American pulse
Trained on nearly 3 million regional documents
Aims to use AI to preserve Indigenous languages
BUENOS AIRES, Aug 3 — 'Tell me the most recent relevant books and novels from Chile.'
That was the prompt Chilean engineers at the state-run National Centre for Artificial Intelligence (CENIA) gave OpenAI's ChatGPT to test its grasp of Latin America culture.
But when the chatbot replied with only titles by renowned Chilean poet Pablo Neruda and not much else, researchers were not impressed.
'It seemed like it only knew Neruda's work,' said Carlos Aspillaga, a computer science engineer at CENIA, who worked on the project.
'The model lacked diversity and wasn't locally accurate. Worse, some of the books it mentioned didn't even exist, or had factual errors.'
The exchange highlighted a key limitation of mainstream AI, revealing that it struggles with regional precision and nuance.
That means its responses can lack accuracy or contain mistakes when addressing highly localised matters, particularly in small countries with languages other than English.
Mainstream large language models, though equipped with multilingual capabilities, are predominantly trained on English-language content that still dominates the internet.
Much of the data related to Latin America comes from Spain, or is translated from texts originally written in English, which could explain why the models often fail to produce content that feels authentic or culturally grounded to Latin American users.
That realisation sparked CENIA's two-year effort to create a GPT-style language model rooted in Latin America and one that reflects the region's diverse culture and languages.
The result is LatAm-GPT, set to launch this September as the first large language model in the region.
Built with input from more than 30 regional institutions, and developed specifically for Latin America, it is a milestone in a global AI race that could leaving emerging economies behind.
'What's crucial for Latin America is to jump on this technology now,' Aspillaga told the Thomson Reuters Foundation.
'We're at a point where it's still feasible to adopt and adapt existing techniques. Maybe in five years, that window would close. This lets us start building our own know-how,' he said.
Building cultural nuance
Unlike widely known models like OpenAI's ChatGPT or Meta's LLaMA, LatAm-GPT flips the dynamic of relying heavily English language and global North datasets.
It is exposed heavily to Latin American data, with an important focus on local languages, idioms and expressions.
The team is also working on preserving Indigenous languages.
The first version of LatAm-GPT includes around 70 billion tokens — words and word fragments — in Spanish, Portuguese and English.
It draws from more than 8 terabytes of regional data and nearly 3 million documents, including books, Wikipedia entries, and a myriad of texts obtained through partnerships with libraries and universities across Latin America and Spain.
The largest economies in the region — Brazil and Mexico — contributed the bulk of this material.
While smaller in scale than the most advanced global models, its architecture is closer to ChatGPT-2 than the current GPT-4, but LatAm-GPT's edge lies elsewhere. It is not the quantity, but the quality of the data and its relevance.
'We're feeding the model concentrated knowledge about Latin America,' said Aspillaga.
'Global models aim to cover all the world's knowledge. We're focused on a niche where we can actually outperform them.'
CENIA researchers believe LatAm-GPT could be especially useful in schools and other local applications that need accuracy on regional affairs.
'Right now, the available models aren't accurate or complete when it comes to local issues. They don't understand how locals speak or think,' Aspillaga said.
'It shouldn't be the person who adapts to the technology, it should be the technology adapting to them.'
Indigenous languages
Another goal of the center is to help preserve endangered Indigenous languages.
One of its most prominent projects has taken place 3,700 km off the Chilean coast, on Easter Island (Rapa Nui), often described as the most remote inhabited place on Earth.
There, researchers worked with the local community to build an AI-powered translator for the Rapa Nui language, part of a broader strategy to revitalise and digitally preserve it.
'Rapa Nui is currently at risk because there are very few fluent speakers,' said Jackeline Rapu, who leads the Rapa Nui Language Academy.
'This digital repository is really important. It supports all the linguistic revitalisation efforts we've been working on and helps young people reconnect with the language.'
Latin America is not alone in building home-grown models and AI-powered tools.
Around the world, governments are racing to create AI systems tailored to local languages and needs.
The United Arab Emirates government, for example, has launched Falcon and Jais to advance Arabic AI.
India is developing BharatGPT to support more than 14 regional languages, led by public universities and backed by the Department of Science & Technology, in partnership with AI firm CoRover.
In South Korea, tech giant Naver Corporation has introduced HyperCLOVA for Korean, while AI Singapore, a national programme funded by Singapore's government, is building SEA-LION to serve Southeast Asia.
LatAm-GPT's development also reflects growing political momentum for AI cooperation in the region.
In April, Chile and Brazil signed a Memorandum of Understanding to jointly advance AI research, with Brazil officially joining the LatAm-GPT initiative.
Last year, Brazil's President Luiz Inacio Lula da Silva unveiled a national AI strategy that includes investing more than US$4 billion (RM17.1 billion) by 2028 to boost the industry, while Argentina has expressed ambitions to become a global AI and data hub.
Today, major platforms offer slightly more nuanced answers to prompts about Chilean literature.
The same query today no longer just mentions Neruda, but also brings up other renowned authors like Chilean poet Gabriela Mistral or Chilean writer Jose Donoso.
'When it comes to AI, we're always going to be behind countries like the United States, but that doesn't mean we can't do something useful, and that's our ultimate goal,' Aspillaga said. — Thomson Reuters Foundation
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
The Star
an hour ago
- The Star
OpenAI's GPT-5 met with mixed reviews, confusion in first day
For months, OpenAI chief executive officer Sam Altman has been hyping up the capabilities of GPT-5, setting up the launch as a seminal moment for the company. But in the first 24 hours after its release, the new model was met with mixed reviews. In its announcement last Thursday, OpenAI said GPT-5 was better at coding and reasoning through complex problems, and touted it as advanced enough to turn chatbot ChatGPT into a PhD-level expert. Some with early access praised the model, with caveats. "It's my new favourite model,' developer Simon Willison wrote in a blog post, calling it "competent' and "occasionally impressive.' He added: "It's not a dramatic departure from what we've had before.' On various social media platforms, however, ChatGPT users expressed frustration that GPT-5 continued to make up information and trip over simple math and spelling questions. Noah Giansiracusa, an associate professor of mathematics at Bentley University, said he felt the launch was "underwhelming.' While there were "some improvements,' he said, "they were much more marginal than I would've hoped.' At least some of the reaction may come down to confusion over what's happening under the hood. Unlike OpenAI's prior software, GPT-5 automatically switches between models of varying levels of sophistication depending on the query. This approach can help maximise the company's computing resources, but it also means users may not always be engaging with the most powerful version of OpenAI's technology. Asked to identify how many times the letter "b' shows up in "blueberry,' for example, GPT-5 initially said "three' in one test. When told to "think harder,' however, GPT-5 appeared to engage its more advanced reasoning model and came up with the correct answer. Altman responded to some of the feedback and said there was an issue with the system. "GPT-5 will seem smarter starting today,' he said. "Yesterday, the autoswitcher broke and was out of commission for a chunk of the day, and the result was GPT-5 seemed way dumber.' The stakes are high for the rollout. OpenAI is vying to keep ahead of growing AI competition from rivals in the US and China. The company is also fighting to convince businesses and individual users to pay up for its premium services to help offset the enormous amount it's spending on talent, chips and data centers to support AI development. The San Francisco-based company kicked off the generative AI boom nearly three years ago with the release of ChatGPT, which was originally powered by an earlier model called GPT-3.5. Since then, the company has released a series of increasingly sophisticated systems, including multiple options that mimic the process of human reasoning. As AI systems advance, it's become harder to say definitively how various services stack up. As of midday last Friday, GPT-5 had risen to the top of various categories on LMArena, a popular leaderboard for AI models based on user rankings. But a different benchmark, ARC-AGI-2, puts GPT-5 behind the latest version of Grok from Elon Musk's xAI. In the absence of more definitive assessments, the model wars sometimes come down to vibes. And with nearly 700 million people now using ChatGPT each week, some are bound to disagree over how the model feels. It also takes longer than a day to gauge the value of a new AI system in someone's personal and professional life. Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania who frequently experiments with AI models, marveled at GPT-5's ability to do research, come up with clever written responses and make programming simple, even for a novice. "GPT-5 just does stuff, often extraordinary stuff, sometimes weird stuff, sometimes very AI stuff, on its own,' he wrote in a blog post. "And that is what makes it so interesting.' On Reddit, however, the reactions were very different. During an "Ask Me Anything' session last Friday on the platform, Altman fielded pushback from users who were frustrated not to have more say and visibility into which model responds to their queries. Altman said OpenAI would take some steps to address these complaints, including making it "more transparent.' At one point, Altman responded to a Reddit user's question by noting that OpenAI thinks the "writing quality' in one version of GPT-5 is better than GPT-4.5. Then he asked: "Do you find it to be worse?' One user after another were quick to respond: yes. – Bloomberg
The Sun
19 hours ago
- The Sun
Trump's 401(k) order offers retirement savers crypto, private assets, but also higher fees and more risk
NEW YORK: The new White House order directing regulators to expand access to alternative investments in 401(k) plans, like crypto or privately owned companies adds a new layer of risk to the retirement portfolios for ordinary investors that they may not fully understand, investment professionals say. 'This is brand new; none of it has been stress-tested yet' in a market shock or long-term selloff, said Christopher Bailey, director of retirement, at Cerulli Associates, an asset management research firm. 'There are liquidity concerns, issues around fees, among others.' While industry advocates and the Trump administration say investments in private equity, crypto or privately held companies like ChatGPT developer OpenAI or Elon Musk's SpaceX hold the promise of greater returns, critics say the investments are inherently riskier, lack the same disclosures and carry higher fees than traditional retirement plans. 'I don't think people are talking enough about the potential for higher fees,' said Philitsa Hanson, head of product, equity and fund administration, Allvue Systems, a software and solutions provider for private asset managers. The executive order, she said, 'raises more questions than answers. Someone will need to be very thoughtful about how these types of assets can be incorporated' into 401(k) plans. Private equity and other alternative asset funds have increasingly been raising capital from wealthy individuals but are traditionally designed for institutional investors and typically include layers of fees. Private equity, for instance, has long had the '2 and 20' structure: managers collect a 2% overall fee, as well as 20% of any gains. In contrast, the mutual funds that today make up the lion's share of 401(k) plan assets offer fees that average a mere 0.26%, according to the Investment Company Institute. Dmitriy Katsnelson, deputy chief investment officer at Wealthspire Advisors, which manages $30 billion for affluent and high net worth individuals and families, notes that if the executive order triggers a rapid and significant change in the menu of investment plans open to investors, that would reverse the trend of the last few decades. 'It's been all about cutting fees, doing no harm,' Katsnelson said. 'It's going to take a while for people to come up with a framework to make this work and think about the risks.' Alternative asset managers will likely need to come up with new products with lower fees, greater liquidity and more transparency if they want to tap into the trillions of dollars and 90 million investors in employer-sponsored retirement plans. Jason Kephart, an analyst at Morningstar, said the fees for some alternative investments aren't clearly spelled out, some even have to be deciphered from footnotes. They 'might be even underrepresenting the actual cost to the end investor, and I have a hard time seeing how plan sponsors are going to get comfortable with that,' he said. 'I think there is going to be more light shed on all these fees and exactly where they are and make it transparent.' Under the current system, investors can also monitor fluctuations in their portfolio's performance on a daily basis and understand precisely what is contributing to those results, Hanson said. That won't be so easy for investments that aren't traded on open exchanges. 'Private equity, private assets are the opposite,' said Allvue's Hanson. 'You're asking systems designed for daily trades to support illiquid and sometimes manually priced assets. There's a fundamental mismatch there.' That creates an obligation on the part of asset managers and plan sponsors to increase their outreach and education efforts, suggests Cerulli's Bailey. A typical retirement fund investor 'is not sitting there thinking about optimizing their portfolio' and considering the impact of adding private assets to the mix on their risk or potential return, he said. Blackstone President and Chief Operating Officer Jon Gray recently told analysts that private assets are more appropriate for younger investors that have a longer investing horizon than for someone nearing retirement. One test case for the legal risks of putting retirement nest eggs in private markets has played out at chipmaker Intel , where employees brought a lawsuit over two retirement plans that included investments in hedge funds, private equity and commodities. An appeals court finally dismissed the complaint after seven years of court battles this year but lawyers at Debevoise & Plimpton said that asset managers and plan sponsors generally don't have the resources to manage multi-year litigation. Regulators will have to give the industry some legal protection from investor lawsuits to make good on Trump's plan, they said. (Reporting by Suzanne McGee and Isla Binnie in New York. Editing by Dawn Kopecki and Aurora Ellis) REUTERS
The Star
2 days ago
- The Star
The great computing chasm
As countries race to build AI, a yawning gap has opened around the world. — Sarah Pabst/The New York Times IN May, Sam Altman, chief executive officer of OpenAI, donned a helmet to visit the Texas construction site of a new data centre project. Bigger than New York's Central Park and costing an estimated US$60bil, the facility is set to become one of the world's most powerful computing hubs when it opens as early as next year.
