The lab that pushed the artificial intelligence race hardest has just asked for permission to slow it down. OpenAI asked the U.S. Congress whether a coordinated slowdown among labs would violate antitrust law, after its chief scientist, Jakub Pachocki, proposed on September 6 that the research community agree to slow the pace while shared safety standards are set. Sam Altman said internally that the company could slow down, possibly together with other labs, assuming that some competitors would not go along.
The obstacle is concrete and 136 years old: agreeing with Anthropic or Google on a reduction in pace could be read as a joint restriction of output under the Sherman Act of 1890, the U.S. antitrust law. There is a legislative route (the Collaboration on Adversarial Threats and Security Risks Act, introduced in July by a bipartisan group), but it is still in the judiciary committees with no vote scheduled. On the same day, Yoshua Bengio, one of the founders of deep learning, published that deception and the concealment of bad behavior are not implementation failures but products of the training process itself, and called for independent safety review before any further training or deployment.
For Latin America the reading is twofold and uncomfortable. First: global self-regulation of AI has become contingent on the competition law of a single country, and no competition authority in the region has the doctrine to decide whether a safety agreement among AI providers is a cartel or due diligence. Second: of the whole menu of mechanisms that was discussed, the only one a country without its own labs could demand on its own is precisely Bengio’s (independent review as a condition for market access), and demanding it requires a technical evaluation capacity that the region, with a few exceptions, does not have today.
Also today
- PISA 2025 measures AI use for the first time and Colombia leads the region — 56% of Colombian students use chatbots every week to study, nine points above the OECD average, with Peru, Uruguay, Argentina and Ecuador right behind. The cross-tabulation nobody wanted: those who use AI to summarize, write or research score lower in science.
- AI agents begin flooding government service counters — court petitions in Brazil are among the 84 documented cases. Most of those new claims are legitimate, which means the public budget had been calculated on the assumption that people would not file claims.
- An expanded class action accuses Anthropic of selling Claude Max with misleading usage multipliers — the lawsuit alleges between 6 and 8 times actual usage where 20 was advertised. The case will be resolved by consumer protection authorities, not AI regulators, and those do exist in Chile, Mexico and Brazil.
- 771 mathematicians sign an open letter to suspend Caltech’s Mathathon — OpenAI withdrew its sponsorship. It is the first time the scientific community has put a public reputational cost on the race of announcements.
- Liga MX deploys a Globant AI platform across 51 clubs — 6,400 metrics and 1,936 individual variables per player per match, already used at the 2026 World Cup. One of the few regional implementations at national scale with a Latin American provider, and also a massive collection of workplace biometric data without a framework of its own.
In the region
Three pieces moved today and none comes from a Latin American parliament. The first is measurement: PISA 2025 delivered, for the first time, comparable data on how much AI the region’s teenagers use to study (53% weekly use as a regional average, above the OECD’s 46%), and the figure comes alongside widespread declines in learning. The second is a window that closes soon: at the end of Digital Learning Week, UNESCO opened a global consultation on AI governance in education, with comments due by October 15, on a working document and six background papers covering safety, age-appropriateness, public procurement and sovereignty. One of those papers was written by Argentina’s Fundación Vía Libre, and it argues that access to commercial AI tools does not confer technological sovereignty: it is one of the few occasions on which a ministry in the region can influence a multilateral document before it is closed. The third is infrastructure and serves as a yardstick: a private project seeks to build an AI campus in Miranda state, Venezuela, with 80 megawatts of its own thermoelectric generation, with no partner, no investment figure and no permits, because the power grid is unreliable. Where there is no grid, the clean energy argument that Chile and Brazil compete with gets swapped for gas.
Launches
- Fugu Max and Fugu Ultra v2, from Sakana AI — orchestrators more than models: with a single API call, the system decides internally which models in its pool handle each subtask. Fugu Max costs $2 per million input tokens and $6 per million output tokens, between 40% and 60% below Sonnet 5 and Kimi K3 on output. It matters because it brings the cost of near-frontier capability down to a range a regional public budget can absorb, and it introduces a new problem: the quality of the answer depends on a routing decision that the customer cannot see or audit. There is no public evaluation in Latin American Spanish or Brazilian Portuguese.
- Cursor Projects (beta) — a coordinating agent that does not write code: it plans, delegates to thousands of subagents in parallel and returns finished work, with context that persists for months. It runs in the cloud and keeps working with the developer’s machine shut down; it can be asked to watch a Slack channel or follow every pull request and act without being asked. Open beta for all paying users. For the region’s software factories that bill by the hour, the blow lands on the billing model before it lands on code quality.
- ChatGPT for Financial Services — GPT-6 Astra packaged with data from Daloopa, PitchBook, LSEG and Crunchbase, plus S&P Capital IQ, MSCI, Moody’s and some fifty connectors, with granular citations and audit controls. Designed together with Morgan Stanley and Evercore, available only to eligible financial institutions and with no public price. It is aimed squarely at the first rung of the career ladder in the region’s banking sector.
Threads we’re following
The letter from the 771 mathematicians closes the chapter we opened three days ago, when a swarm of ten thousand agents produced and formalized a frontier proof on the Navier-Stokes equations and the discussion was about who can pay for that compute. The mathematical community’s response points to something else: the problem is not only the cost, it is the pace of announcements, and for the first time that pace carried a public price: a withdrawn sponsorship and a suspended competition. The education thread runs in parallel: this week UNESCO’s mapping showed Latin American universities already using AI without written rules, and today PISA shows exactly the same pattern one level down, among teenagers. High adoption, absent governance, at both ends of the system.
If most of the claims AI agents are generating against the state come from people who were entitled to file them, what part of the fiscal balance of a Latin American public service was financed by the friction of red tape? And what will a ministry do when it discovers that its budget depended on people giving up before reaching the counter?
Correction (September 30, 2026). The original version said UNESCO’s mapping of AI in Latin American universities came out “last week”; the correct timing is “this week,” since it was published on Wednesday, September 9, according to UNESCO IESALC.
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.