Terence Tao, one of the world’s most renowned mathematicians, described how modern coding agents (AI programs capable of writing and reorganizing software with almost no human intervention) took mathematical projects that had been abandoned for 27 years and migrated and modernized them in a matter of hours. In the process they generated visualizations that used to take weeks and detected two errors that Tao himself had never seen. His account, published in Old and new apps, via modern coding agents, went viral on July 12.
What is interesting is not that AI writes code (we already knew that), but that here it acts as a real research collaborator: it does not just produce, it also finds flaws an expert missed. And these tools are increasingly cheap and accessible. That raises a question with a Latin American edge: if a researcher no longer needs a large team or their own supercomputing to recover and audit years of work, perhaps the region’s universities, which almost never had that access, can compete in a different way. The other side is less friendly: the more we depend on these agents, the more we depend on infrastructure that is built and decided outside the region.
That contrast shows up in the ecosystem’s own numbers. AI-native startups are growing 145% a year in Latin America, a remarkable pace just as the capital gap with the big labs is becoming structural. Growing fast and depending on other people’s parts are not incompatible; they coexist, and that tension defines much of the moment.
Also today
- Anthropic extends free access to Claude Fable 5 until July 19 — It is the third time in five weeks that it has postponed usage-based charges. Lack of capacity or managing expectations ahead of an IPO? For now, more free time for anyone who wants to try it.
- Cursor is preparing “Sand,” an office agent for non-programmers — It aims to compete with AI-assisted work tools, although its launch depends on how SpaceX’s purchase of Cursor progresses.
- Musk calls Sam Altman “Scam Altman” — Apple’s lawsuit against OpenAI became a new front in the long personal rivalry between the two.
In the region
Beyond startups, the region is making moves on infrastructure and policy. Argentina is assessing its potential as a host for AI data centers, betting on cheap energy and the Súper RIGI incentive regime, although its power grid is not yet ready to scale. On the risk front, a report claims that AI-driven identity fraud in Latin America is the highest in the world, at 48.3%; it is a striking figure but from a single source, so it is best taken as a warning sign and not as settled data. In public policy, Mexico is discussing incorporating AI into the curricula of its New Mexican School (Nueva Escuela Mexicana), still without a committee report, and Panama formalized an Artificial Intelligence Subcommittee within its new critical technologies commission. As a backdrop, a column in Infobae argues that the region is more exposed than Europe to an eventual “blackout” of models decided in Washington, precisely because it lacks a comparable ecosystem of its own.
Threads we’re following
At OpenAI, we are closely following the internal reorganization in the midst of preparations to go public. This week brought the departure of Johannes Heidecke, head of safety systems since 2024, after a restructuring that merges research and safety under a single leadership. It comes days after an independent index gave the company a mediocre safety grade. It is one more chapter of a story we have been following: what happens to the safety teams of the big labs when commercial and stock market pressure tightens.
If the tools that level the playing field are getting cheaper, but infrastructure and access rules are still decided elsewhere, is growing fast enough, or is the real bottleneck something else?
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.