Culture as corpus

The team at Chile's National Center for Artificial Intelligence released the first LatamGPT model. It started from Llama 3.1, a base model with 70,000 million parameters, applied continued pretraining with a Latin American corpus, tuned the result with human supervision, and published it as open access.
With the data cutoff of 31 July 2025, a window closes. What had gone in up to then —188 million documents, 296,500 million tokens, 152 datasets from twenty countries— became the raw material of a language model deliberately trained on the memory of a region.
Fifteen people produced the artifact. The telling figure of the project lies elsewhere: of 429 websites examined for inclusion, 125 survived. Less than a third of the candidate material made it through the licensing audit.
That survival rate works as a thermometer of something broader than a copyright problem. It describes the resistance that the Latin American institutional environment puts up against the flow of its own information. The material exists, the willingness to share it exists, and the capacity to process it exists; the gradient that would allow it to move from the archives to the computing systems remains blocked by a layer of legal indeterminacy that each project must dissolve on its own, site by site, with manual review and a lawyer embedded in the technical pipeline.
Europe resolved that friction, for example, with the text and data mining exception; Japan and Singapore did so with provisions on computational analysis. In Latin America the cost of ambiguity is paid in hours of human labor and in lost material.
Building on someone else's land
LatamGPT is built on Llama, a base model released by Meta under terms that the company defines and can revise. That dependence is the project's founding tension and deserves to be named precisely. The region contributed the culture, the curation, the governance, and the labor; the computational substrate on which all of that rests belongs to a corporation operating under a logic of perpetual lease. The architecture recalls feudal arrangements in which the vassal's labor improves land whose ownership never changes hands. The improvement is real, the knowledge acquired is real, and the starting condition remains.
Acknowledging that clearly is worth more than ignoring it. Training a seventy-billion-parameter model from scratch requires a computing infrastructure that no Latin American country has on its own, and postponing the project until that infrastructure exists would have meant giving up an entire generation of learning. The team chose to move forward on the available substrate while accumulating the only thing the region can accumulate with its own resources: curation capacity, legal judgment, networks of donor institutions, and operational experience handling cultural data at scale. That is the capital that later makes it possible to negotiate from a different position.
The process is a form of symbolic industrialization. Document collections, oral records, academic output, and everyday conversation enter a transformation chain —heuristic cleaning, hash-based deduplication, multidimensional classification, syntactic validation— and come out converted into computable tokens. Culture goes in as heritage and comes out as an industrial input. Every transformation of this kind produces waste: what is not digitized, what has no clear license, what nobody cataloged in time. The languages and communities left out of the corpus do not vanish from the world, but they do vanish from the space of what machines can represent, which in the coming decades will be a growing portion of the space of what can be collectively thought.
The license as an instrument of policy
Technical artifacts embed political decisions in their material structure. LatamGPT offers a clean illustration of that thesis. The project set up three levels of corpus availability: open access on Hugging Face, access on formal request for research, and exclusive use under an institutional agreement. On top of that gradation it added a clause that bars entities with annual revenues above one billion dollars from using the material to train, fine-tune, or evaluate artificial intelligence systems without prior written authorization.
That clause is a political decision encapsulated in a license. It declares that Latin American cultural knowledge, aggregated and curated with public funds, is not raw material for free extraction by those who already concentrate the capacity to turn it into product. The wording is open to debate —the threshold is arbitrary, enforcement will be difficult, the original third-party licenses remain intact—, and its existence marks a shift in how the region conceives of its own informational heritage. For twenty years Latin America handed over the record of its digital life for free to platforms that capitalized on it without compensation. Now an instrument appears that discriminates between uses and users.
The unresolved tension lies at the other end of the chain. The Trueque benchmark gathered 2,296 questions from twenty countries through voluntary contributions from more than a hundred institutions and from individuals who wrote, from their situated knowledge, what a Nicaraguan almíbar is or what causeo is prepared for in Chile. That work, which no machine could produce and which no corpus extracted from the web contained, was done without pay. The architecture of reciprocity that would compensate those who donate the knowledge that gives the system its value has yet to be designed, and its absence reproduces on a small scale the pattern that the billion-dollar clause tries to correct on a large one.
Knowledge that does not travel in documents
Know-how resides in networks of people and resists being codified in text. An economy does not acquire productive capacity by importing manuals; it acquires it by hosting those who already know how to do things and letting them do them. That observation explains why the distance between Chile and the rest of the region in the Latin American Artificial Intelligence Index —70.6 points against an average close to 41— owes less to budget differences than to differences in accumulated practice.
Chile learned how to govern a multinational corpus while governing it. It discovered that legal counsel had to be integrated into the engineering team, that access levels had to be defined before collection, that human validation could not be delegated entirely to automatic filters. None of those lessons appeared in its national artificial intelligence strategy, because none can be derived by deduction. They emerged from contact with the material. Fifteen people and twelve months of work generated a stock of institutional competence that now resides in Chilean heads and will be transferred, in part, to those who take part in the next iteration. That is the deeper reason why experimenting is public policy and drafting principles is preparation for public policy. Countries that build accumulate the kind of knowledge that is deposited in people and reactivated in the next project. Countries that observe accumulate documents.
The cycle that is closing
LatamGPT 2.0 states its objectives: to deepen regional knowledge by incorporating dialects and Indigenous languages —Quechua, Guaraní, Aymara, Nahuatl, K'iche', Q'eqchi', Mapudungun, Tikuna, Wapishana—, to add reasoning and agentic behavior capabilities, and to expand the network of alliances. The Dominican Republic ranks fourth among the countries that contributed the most questions to the benchmark, a sign of an awake technical community, and appears with two institutional agreements against Chile's thirty and Argentina's ten. Individual participation far exceeded organizational commitment.
Correcting that asymmetry does not depend on computing infrastructure or legislative reform. It depends on universities, archives, media outlets, and public entities with document holdings determining what they own and under what license they can share it, and making contact with a project that explicitly states its interest in underrepresented countries. The channel is open and the window for the second version remains active.
The systems being trained today will define for decades which portions of Latin American experience are legible to the machines that will mediate access to knowledge. The composition of those corpora is being fixed now, in minor administrative decisions that no assembly will debate and no headline will record. The region has a short interval ahead to determine how much of itself will be inside.
Culture disappears from the public imagination as long as it does not connect with the digital flows that are accelerating our economy.