A swarm of some ten thousand OpenAI agents worked 88 hours straight on one of the seven Millennium Prize Problems and produced a proof. According to the company’s announcement, the agents were deployed on August 28 on an internal model that has not yet been released, and the result is that the Navier-Stokes equations (the ones that describe how fluids move) can break down: an initially smooth fluid at rest, pushed by a smooth force, reaches a singularity in finite time. It pays to be precise, because much of the coverage mixed things up: what was proven were statements C and D of the Clay Mathematics Institute’s official formulation, not global smoothness, and OpenAI stated that it will not claim the $1 million prize.
What makes this more than a corporate announcement is that the proof comes formalized in Lean, a language that lets mathematics be written so that a computer checks every step. Its correctness does not depend on believing the model: anyone can verify it. And that is where the figure that matters for the region appears. Lean is free and open, and Latin America’s mathematical human capital is real and competitive. What is neither free nor open is the engine: 130 billion output tokens and 2.7 million messages between agents for a single problem. A mathematics department in Santiago, São Paulo or Mexico City can audit this result tomorrow morning; none of them can produce it. The barrier to entry for frontier mathematics is no longer knowledge but the inference bill.
There is also a dispute worth following. The mathematician Tristan Buckmaster, who together with Levent Alpöge published work on key subproblems days before the announcement, accuses OpenAI of building on that advance; the company denies it and the chronology remains unresolved. It is the first time the question of academic credit has come up when the human input goes into a private system that no one can inspect.
Also today
- NSA, CISA and FBI accuse six Chinese companies of distilling Anthropic, OpenAI, Google and xAI models on an industrial scale — DeepSeek, Moonshot, Alibaba, MiniMax, StepFun and Z.AI are, almost exactly, the list of open models the region builds on. If distillation (training a model of your own from another model’s answers) becomes prosecutable, Beijing will not be the only one paying the cost.
- Cognition raises more than $2 billion at a valuation of $48 billion — Annualized revenue for its programming agent Devin went from 492 million to nearly 900 million in four months, and its reference customers are banks and automakers, not startups.
- Qualcomm and Amazon sign a deal for up to $60 billion in inference silicon — Qualcomm’s first contract with a major Western cloud provider, with warrants worth 4 billion. The hardware battle is shifting from training toward inference: the same cost that today separates auditing from producing.
- Anthropic leaves the Information Technology Industry Council over chip export controls — The trade group asked for three chip control provisions to be removed from the annual U.S. defense package, and the company walked out.
- Chicago’s first school without teachers opens: two hours of AI a day and $55,000 in tuition — Two days after the largest public school districts in the United States banned AI for their students. The dividing line is not who uses it, but who buys it.
- A Fields medalist founds an institute to measure AI risk and joins OpenAI the same month — Jacob Tsimerman’s institute starts in January 2027 with thirty mathematicians, sixteen months behind the capability it aims to measure.
In the region
After three days without AI news originating in the region, two stories appear, and one of them carries weight. On September 6, Mexico’s SECIHTI reported that Mexico and China agreed on a new stage of scientific cooperation focused on artificial intelligence and supercomputing, signing a memorandum of understanding that enables joint research between universities and companies, exchange of policy frameworks, public-private partnerships for digital infrastructure and transfer of advanced technology. It is the first time since the binational subcommission was created in 2004 that the Chinese minister for the portfolio has taken part. The reading cannot be done without the calendar: the country with the greatest trade exposure to the United States puts in writing a path of technological cooperation with the opposing bloc forty-eight hours before three U.S. agencies formally accuse the Chinese AI industry of systematic appropriation, and in the year the USMCA is under review. The second piece of news is about the ecosystem: the 11th Latin American Congress on AI, Technology and Business is being held today and tomorrow in Santiago with more than five thousand attendees from more than fifty countries, framed by a figure worth reading calmly: the regional inventory of colocation data centers grew twenty percent in one year, with demand concentrated in Brazil, Mexico, Chile, Colombia and Peru, a map that matches that of available energy and not that of national AI strategies. The region is building capacity to host compute, which is not the same as having it. And tomorrow, September 10, the first compliance deadline under the Peruvian regulation of Law 31814 expires: health, education, justice, security and finance become obligated to explain what their algorithm decides.
Launches
- ChatGPT Images 2.5 — OpenAI’s image model with up to 50% lower latency, editing by sketches and by natural-language comments, and two new models available to developers. No geographic restriction stated. It matters for an uncomfortable reason: the ability to preserve a person’s identity across chained edits is exactly what underpins the problem of deepfakes and gender-based digital violence that the region already has in the courts.
- AlphaGenome Atlas — Google DeepMind precomputed the molecular effects of the 9 billion possible single-letter variants in the human genome and published the result: a petabyte of data, more than thirty times the AlphaFold database. Website and interface free for academic research. For once, the compute asymmetry is resolved in the region’s favor: the spending was done once and the result is given away, so a genetics lab in the region can query the same thing as one in the North without needing its own servers. The limit is one of data, not access: the genomic training banks underrepresent Latin American and Afro-descendant ancestry, and the authors themselves warn that this is never sufficient clinical evidence.
- Muse — Meta’s personal agent that sends emails, books trips and buys tickets, with a second agent that approves everything that goes out to the internet. A free tier plus subscriptions of $20 and $100 a month. It runs inside WhatsApp, but is being rolled out only in the United States: the channel is regional and the product is not.
Threads we’re following
Three days ago we reported that an Anthropic model had formalized Fermat’s Last Theorem in eleven days, and the point then was the same as today: formalization in Lean turns a claim into something anyone can check without asking anyone for permission or trust. Today’s news adds a different and harder chapter. With Fermat, the task was to rewrite in verifiable language something the mathematical community already considered proven. With Navier-Stokes there is a new result, and it comes with two labels together that we had not seen attached until now: open and free verification on one side, closed and extremely expensive production on the other. This is the same week in which a Fields medalist founds an institute to measure that capability and goes to work for the company that builds it.
If verifying frontier mathematics is free and producing it costs more than any public budget in the region can reach, and at the same time the cheapest shortcut (building on open models) is starting to come under legal suspicion, is accepting the role of auditor an honest strategy or just a dignified name for dependence?
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.