Today’s news is a big claim: OpenAI published a paper arguing that its GPT-5.6 Sol Ultra model, coordinating 64 subagents in parallel, produced in under an hour a proof of the Cycle Double Cover Conjecture, a graph theory problem open since the 1970s (OpenAI document). This adds to the story we had been following this week about coding agents turned research collaborators: on the same day, mathematician Terence Tao published a new visualizer built with the same agent workflow he had documented two days earlier, a sign that it was not an isolated case.
The claim is worth reading calmly. Mathematician Thomas Bloom, of the University of Manchester, called the proof “very nice,” but warned that it omits citations to prior work from 1983 and that there is still no full peer review. In other words, there is a result that impresses specialists and, at the same time, the usual signs that mathematical validation is only just beginning. The distinction matters because it marks the difference between “an AI solved an open problem” and “an AI proposed a proof that the community still has to verify.”
For the region, the angle is not so much the graph problem as access. If a model can orchestrate dozens of subagents to tackle in an hour a conjecture that resisted for half a century, the question is whether that capability, increasingly available by subscription, can level the playing field for researchers at Latin American universities without their own supercomputing, or whether the “discovery” is still captured by whoever controls the model. In the same week that Anthropic starts localizing Claude’s pricing for India, its biggest market after the United States, that question about price and access stops being abstract.
Also today
- Helsing raises $1.8 billion, the largest round for a defense startup in Europe — The German military AI company is now valued at $18 billion, a sign of investor appetite for a sector that barely exists in Latin America.
- Xi Jinping will personally deliver the opening speech at Shanghai’s World AI Conference — Beijing raises the political profile of its strategy just as rumors circulate of new restrictions on the export of Chinese models.
- TSMC reports sales up 36% despite a sell-off in semiconductor stocks — Spending on compute remains firm even as investors get nervous.
- Half of users do not know what AI can do beyond a basic search — According to a Stagwell/NRG study, the gap between basic adoption and sophisticated use is wider than market enthusiasm suggests.
In the region
The week brought two concrete regulatory moves. Mexico confirmed that its Congress will debate an AI legal framework starting July 19, now with uncertainty over the USMCA pushing toward an approach that is more domestic than harmonized with the United States and Canada. And Ecuador presented the region’s first bill focused specifically on the use of AI within the judicial system, with disciplinary sanctions for anyone who uses AI-generated content without verifying it. In Chile, the Senate is still reviewing its risk-tiered regulation bill, with no substantive news this week. Added to this is a striking precedent: Canada’s banking regulator became the first national financial supervisor to cite an AI model by name in cyber risk guidance aimed at banks, something the region’s financial superintendencies could start to replicate.
Launches
- Perplexity Computer, update — It adds memory, live model switching, publishing of your own websites and financial data on private companies, and brings back Claude Fable 5 as the orchestrator for long tasks. Available to Pro and Max users.
Threads we’re following
This is the third time this week that the idea of “AI agents as research collaborators” has appeared in the news. It began with Terence Tao showing how a coding agent modernized in hours mathematical software abandoned for almost three decades; it continued with a second experiment of his that confirmed the method is repeatable; and today it escalates to a much more ambitious claim from OpenAI. The pattern is clear, although its verification, case by case, is only beginning. It is worth following without yet buying the most optimistic version or dismissing it outright.
If the ability to orchestrate dozens of agents to tackle hard problems becomes a subscription product, will the next big breakthrough come from whoever has the best question, or from whoever can pay for the best model?
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.