The letter urging Washington not to impose “premature restrictions” on open-weights models doubled its list of signatories in just over twenty-four hours: it went from twenty-five companies to about fifty, and OpenAI and Google, two of the big closed-model labs that were missing yesterday, added their names. OpenAI did so after its absence was publicly remarked on X. Open weights are a model’s published parameters: anyone can download them, run them on their own infrastructure and adapt them, without paying for each query.
By the end of the day two were still absent, and not by chance: Anthropic, whose chief executive has called open source “a red herring,” and Amazon, its main investor and the supplier of the Trainium silicon on which Anthropic trains and serves its models. Neither explained its decision. With that, the dividing line stopped being open versus closed and became something else: those who sell compute versus those who sell access. Nvidia, which leads the letter, wins with every open model someone runs on its chips; Anthropic wins with every task someone solves by paying for its programming interface.
For Latin America, the problem is that neither side includes it. The region depends on open weights to run models on its own servers, maintain sovereignty over its data and sustain projects such as Latam-GPT, and it appears neither among the fifty signatories nor at the table where the decision is being settled. The scale of the imbalance became clear the same day: SK Group and Nvidia expanded their alliance by more than $500 billion, with a 2-gigawatt artificial intelligence factory in South Korea by 2027 and a long-term memory agreement with SK hynix, signed with the South Korean president present. Chile, which leads investment in the region, will devote about $1 billion to artificial intelligence this year. The bottleneck that contrast exposes is not talent: it is memory and energy.
Also today
- Monday.com becomes the 21st company to attribute layoffs to AI in 2026 — about 620 people, 20% of its workforce, with the now-canonical formula: “it was not to cut costs or to replace people with AI.” The annual tally stands at about 140,000 tech jobs cut in the United States, and none of the 21 companies publishes a breakdown by country.
- Fifteen models tried to fly a drone to find and follow a person in an office — the best one outperformed the human-AI teams in four of five subtasks. The obstacle still standing is not following the person: it is mentally reconstructing the room.
- Brazil’s data agency publishes the first report from its AI regulatory lab — since March, the ANPD has been working with three companies (Metatext, Synapse AI and Prevvine Tecnologia) that are testing AI systems in a supervised environment. Two useful findings for any country copying the model: coordination among participants was the first point of friction, and the synthetic data protocol proved insufficient.
In the region
The regional move of the week comes from Chile, and from Congress rather than the executive branch: the Chamber of Deputies approved in general terms the bill that penalizes realistic AI-generated digital imitations of a person’s image, body and voice (Boletín 17.795-19), with 128 votes in favor, fines of 5,000 to 10,000 UTM (up to about 716 million Chilean pesos) and the recognition of a new “right to digital integrity.” It returns to committee for amendments and still has to go through the Senate, in a country that does not yet have a general AI law or a fully established data authority to enforce fines of that size. The bill adds to a regional pattern with three different ways of tackling the same problem: Chile penalizes the use, Brazil banned the tool by decree in May, and the European Union has banned it by regulation since December. None of the three solves the case where the generator is hosted outside the jurisdiction. And at the back of the calendar is the date that orders everything else: on August 2, the transparency obligations of Article 50 of the European AI Act take effect, with final guidelines adopted on July 20, and several Latin American frameworks under consideration are modeled on that text.
Threads we’re following
We had been following the fragility of access: how long frontier tools last in the region when the decision to switch them on or off is made in another capital, after a government shutdown mechanism left Latin America without access to a cutting-edge model within hours and without warning. Today’s chapter does not change the board, it clarifies it: the fallback route for that scenario (downloading the weights and running the model yourself) now has fifty companies defending it in Washington and two defending the opposite with the same right to be in the room. The region, which would need that route most, still has no voice on either side.
If on a single day a business letter in Washington defines the region’s future access to open models, and a two-gigawatt factory is signed with a head of state present, what is the equivalent instrument with which Latin America negotiates? And if the answer is a ministerial declaration, are we playing the same game or watching it being played?
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.