The week closed with a governance decision that redraws the map of artificial intelligence: the US government permanently scrapped the only oversight mechanism that had been proposed for the most advanced models, a voluntary 90-day safety review scheme before each launch. It was not postponed: it was discarded. The executive order would have created a first public checkpoint for technologies that today operate without any.
What stands out is the contrast with what those same models are achieving. This week OpenAI announced that a general-purpose model disproved the Erdős conjecture, a problem that had been open for 80 years, without having been trained specifically for mathematics. Mathematician Tim Gowers, a Fields medalist, described it as a “milestone in mathematical AI.” Capabilities that grow at full speed and oversight that disappears are, for now, moving in opposite directions.
For Latin America the scene has an uncomfortable side: the region has no seat in these decisions and no market weight to influence them. For reference, the region’s most ambitious sovereign compute budget (the Coatlicue project in Mexico, about 6 billion pesos over two years) is equivalent to about 1% of the funding round Anthropic is about to close. The underlying question is what the region’s governments can negotiate when they control neither the benchmark regulation nor the money.
Also today
- An Anthropic co-founder predicts an AI-assisted Nobel within 12 months — Jack Clark said at a conference in Oxford that an AI will collaborate on a Nobel-level discovery before May 2027, and reaffirmed that existential risk “has not gone away.” He says it just as his company closes one of the largest rounds in history.
- Anthropic is on track to close more than 30 billion at a valuation above 900 billion — According to Bloomberg, Sequoia, Dragoneer, Altimeter and Greenoaks are leading. If it goes through, it would overtake OpenAI as the most valuable private startup in the world.
- Anthropic buys Stainless for more than 300 million — It takes control of the infrastructure that generated software development kits (SDKs) used by OpenAI, Google and Cloudflare. It is shutting down the hosted products, although customers keep the SDKs already generated.
- Meta tracked its employees’ emails, keystrokes and code to train AI, with no opt-out — Leaked audio from an internal meeting reveals the tracking; the chief technology officer admitted that “there is no opt-out on your corporate laptop.” The audio became public the same day the company announced about 8,000 layoffs.
In the region
The change in the United States moves its stance from “no progress” to “no oversight by design,” and that matters because its framework is usually the regulatory reference for the region. As a counterpoint, Latin America has its own milestone this week: on May 27 Brazil’s Chamber of Deputies is scheduled to hold the floor vote on PL 2338, the region’s first AI regulatory framework with full national scope. It is probably the most relevant AI policy event in Latin America in 2026. The rapporteur’s text has already incorporated government proposals, although tension persists with more than 50 organizations from the creative sector that are asking to keep the copyright chapter.
In parallel, an ecosystem signal: WideLabs, a Brazilian company, is building language models for the governments of Chile, Mexico and El Salvador, adapted to the languages and local contexts of the public sector. The fine print is that it relies on Oracle infrastructure and Nvidia chips, both US suppliers: a “sovereignty” that remains anchored in foreign hardware and cloud, and that is worth comparing with initiatives such as LatamGPT.
Launches
- Gemini Spark (beta) — A proactive agent that runs 24/7 in the background, without needing to open the app, and manages Gmail, Calendar and third-party apps such as Canva, Instacart or OpenTable. It reaches trusted testers this week and, starting May 25, Google AI Ultra subscribers (United States, $100 a month). For now it is only in English and in the US, but with Android on more than 80% of the region’s devices, its eventual arrival would be one of the largest consumer agentic AI deployments in Latin America’s history.
Threads we’re following
The Erdős result and the Nobel prediction add to something we had been observing: a streak in which AI begins to behave like a scientific agent, generating new knowledge instead of just solving competition tests. This week’s news is not a faster model, but one that makes an original contribution to a field. It is the kind of leap that makes the conversation about who oversees these tools and who has access to them more urgent, not less.
If the most capable models advance faster than the rules that should accompany them, what can a Latin American government offer in exchange for a seat at the table?
Correction (September 30, 2026). The original version said the Coatlicue project cost about $6 billion over five years; the correct figure is about 6 billion pesos (about $326 million) over two years of construction, according to El Financiero. At that figure, the project is equivalent to about 1% of Anthropic’s round, not less than 1% as the original version said.
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.