Experimenting with rigor

On 31 July 2025, the data cutoff for the first LatamGPT corpus was closed. What remained inside is 188 million documents, 296,500 million tokens and 152 datasets from twenty countries. On that base, a team coordinated from Chile's National Center for Artificial Intelligence carried out continued pretraining on Llama 3.1 with 70,000 million parameters, applied supervised fine-tuning and published the resulting model in open access. The region produced its first language model deliberately trained on its own cultural memory.
The figure that deserves the attention of Dominican decision-makers appears elsewhere in the project: of 429 websites analyzed for inclusion in the corpus, barely 125 survived the license audit. Less than a third.
That figure precisely describes the state of Latin America's information infrastructure. The material exists, the institutions willing to share it exist, and the technical capacity to process it exists. What is scarce is legal clarity about what can be used, under what conditions and with what traceability.
The LatamGPT team solved the problem in the most expensive way available: fifteen people, successive iterations of manual review backed by site-by-site documentation, verification of usage policies and risk-management filtering. A lawyer embedded in the technical pipeline. That effort produced a reusable asset, and it also made visible that no country in the region currently has an exception regime for text and data mining for research purposes comparable to article 4 of the European copyright directive or to the Singapore and Japan provisions on computational analysis. The absence of that framework turns every regional AI project into a legal negotiation from scratch.
The cost of waiting for the technology to mature
The predominant institutional response to artificial intelligence in Latin America has been to formulate national strategies, create commissions and draft ethical principles. The Dominican Republic moved in that direction with its National Artificial Intelligence Strategy and ranks ninth in the Latin American Artificial Intelligence Index with a score of about 45 points, above the regional average and at a considerable distance from Chile, which leads with 70.6. The gap between the two positions is not explained mainly by budget differences. It is explained by the fact that Chile decided to build something and to learn from the process of building it.
LatamGPT works as an empirical demonstration of that difference. The project did not wait to have data governance settled before starting; it found out how to govern the data while collecting it. It established three levels of availability —open on Hugging Face, access by formal request for research, and exclusive use under institutional agreement— because working with real sources demanded it. It incorporated a restriction that prevents entities with annual revenue above one billion dollars from using the corpus without prior written authorization, a decision that no national strategy document would have anticipated without the prior experience of having compiled the material. The governance architecture emerged from practice, and that is why it withstands contact with practice.
A concept operates here that we could describe as public entrepreneurship: state and academic institutions taking on the risk of building infrastructure that the market does not produce because the private return is insufficient compared to the social return. Chile's Ministry of Science, CENIA, the Data Observatory and the Development Bank of Latin America took on that risk. The result is a regional public good from which the Dominican Republic already benefits without having contributed proportionally to its creation.
The Dominican contribution exists and is uneven
The Trueque benchmark, the project's evaluation component, gathered 2,296 questions from twenty countries through contributions from more than one hundred institutions. The Dominican Republic ranks fourth by volume of questions contributed, behind Ecuador, Colombia and Mexico. That performance reflects an active technical and academic community, willing to participate when a concrete channel for participation is opened to it. The asymmetry appears in the other dimension of the project. The geographic distribution of strategic alliances signed between 2023 and 2026 shows thirty agreements in Chile, ten in Argentina, five in Mexico, Colombia and Peru, and two in the Dominican Republic. Individual Dominican participation far exceeded institutional participation. Individuals contributed; organizations mostly stayed on the sidelines.
That distance between individual capacity and institutional commitment is the most tractable public policy problem among those raised here. Correcting it requires neither a computing infrastructure budget nor legislative reform. It requires that universities, archives, media outlets, public-sector entities with document collections and civil society organizations identify which collections they hold, determine under what license they can share them and establish contact.
For those interested, the address latam-gpt@cenia.cl is published, and the project explicitly states its interest in filling the gaps of the countries with the least represented data.
What the second version makes possible
LatamGPT 2.0 sets three objectives: deepen Latin American knowledge by incorporating dialects and indigenous languages, add reasoning capabilities, tool use and agentic behavior, and expand the coverage of alliances. The list of languages sought includes Quechua, Guaraní, Aymara, Nahuatl, K'iche', Q'eqchi', Mapudungun, Tikuna and Wapishana. The absence of the Spanish-speaking Caribbean from that inventory has a historical explanation, and it also leaves open the question of what the Dominican Republic contributes to regional linguistic representation beyond Dominican Spanish and its registers.
There is a concrete operational window. The project needs collective human validation of the benchmark questions, a task that a Dominican university could coordinate with modest resources and in short timeframes. It needs datasets from specific domains that Dominican institutions hold without ever having cataloged them as training inputs: historical records, academic output, regulatory documentation, press archives. And it needs use cases that demonstrate the model's usefulness in real institutional contexts, an area where the Dominican public sector accumulates document management problems that an open-access, locally deployed model could address without the data transfer restrictions that foreign providers impose.
Applied experimentation produces institutional learning that no strategy document generates. Chile showed it by building a model with fifteen people and a twelve-month data cutoff. The learning acquired in that process —about licenses, about curation, about the real economics of training with one's own data— now belongs to Chile, and is passed on to those who take part in the next iteration. The region will enter the next decade with countries that learned by doing and countries that learned by reading what others did.
The distance between the two groups is being set in this cycle, with decisions that cost less than any national plan and yield considerably more.