Today's Headlines

  • OpenAI is working with an independent mathematics advisory group, and says the same internal model has resolved more than 100 open problems
  • Apptopia estimates that Meta's Muse beat ChatGPT's first twelve days on mobile, matched on iOS and the US and Canada
  • NVIDIA launches DSX Ready, starting with battery energy storage systems and cooling distribution units
  • OpenAI argues the United States should lead an effort to build global technical standards for frontier AI
  • Google details the built-in intelligence in its Googlebook laptop, on shelves in the US on October 4

Today's three stories run along one line: what gets built, and who gets a say in it. OpenAI has tied itself to an outside group that will advise on how mathematical results are released, Meta's new app is ahead on the measure of which app people open, and NVIDIA has made power and cooling — the physical constraints themselves — the object of a qualification program.

Where the people with a say are standing is part of today's reading. The mathematics advisory group sits at the Institute for Advanced Study in Princeton, its members take no payment for the work, and the group says it will publish its recommendations on its own site.

Today's Top Three

OpenAI teams with an independent math advisory group to coordinate how results are released

OpenAI announced that it is working with an independent advisory group of nine mathematicians.

The name and the address are in the announcement. The group is called the Advisory Group on Mathematics and Artificial Intelligence and is hosted at the Institute for Advanced Study in Princeton, New Jersey. OpenAI writes that the group "will serve as a bridge to the mathematical community and broader public, giving mathematicians a voice in how we move forward."

The announcement rests on OpenAI's own figures. The company writes that it began training a new internal model on August 28, and that beyond resolving the Navier-Stokes Millennium Prize problem, the model has now resolved more than 100 long-standing open problems across most areas of mathematics. The pace of that progress surprised the mathematicians inside OpenAI, the company adds, and led to internal discussions about the best way to inform the community so the field can prepare and adapt.

Mathematicians put out a document of their own this month. Twenty-five Fields Medal winners signed an open letter titled "A Severe Misalignment of AI in Mathematics," arguing that AI labs are threatening their intellectual work as they seek to one-up each other with solutions to famous math problems. OpenAI cites the letter in its own post, writing that the mathematicians raise concerns about the negative externalities of solving open problems as a benchmark for new AI systems.

The group has described its founding in its own words. In a guest post on the mathematician Terence Tao's blog, the group writes that its purpose is to advise AI companies on their interactions with mathematical research and with the mathematical community, including the responsible presentation and release of mathematical results, and that it operates independently of any AI company with members who take no payment for the work.

The origin story sits in the same post. OpenAI approached some of the members about establishing an external advisory board, and in agreement with OpenAI they decided to create an independent group and invite others to join.

The group has also named its current task. It says the specific challenge in front of it is advising OpenAI on how to coordinate the release of a large number of significant results in mathematics that OpenAI reports were produced by its internal model, and it has opened a form asking the mathematical community for input on that question.

OpenAI spelled out how far the advice reaches. The group is free to offer advice OpenAI has not requested, to comment on OpenAI's impact on mathematics, to make its advice public, and to change its own membership, while advice on how OpenAI paces its internal progress in mathematics sits outside the group's remit. The Institute for Advanced Study made the same point in its press release: "Although we will give advice, we do not have decision making power at any AI company, and the responsibility for the decisions made by any company will rest with that company."

The overlap between the group and the letter is one person. Of the nine initial members, only the IAS's Camillo De Lellis also signed the Fields Medalists' letter, TechCrunch reports.

The Clay Mathematics Institute review this paper covered on September 13 sits on the other side of the same fault line. Clay's rules make peer-reviewed publication, two years in print and general acceptance in the mathematical community part of the condition for an award, and with no channel for direct submissions, every marker of that acceptance is an act by a third party.

Meta's Muse outpaces ChatGPT's early mobile launch in a like-for-like estimate

Meta's AI app Muse was downloaded more times in its first 12 days on the market than ChatGPT was in the 12 days after its mobile debut, according to estimates from the market intelligence provider Apptopia.

The terms of the comparison come from how the two apps shipped. ChatGPT arrived on mobile worldwide but on iOS only, while Muse is on both the App Store and Google Play and is limited to the US and Canada. Apptopia therefore looked only at iOS data for the US and Canada for both apps, across the same first 12 days of each launch.

That subset carries today's number. Muse has seen 1.8 million downloads against ChatGPT's 1.3 million, Apptopia says. Counted on its own, Muse has reached 2.8 million total installs globally in its first 12 days.

Daily active users show the same gap. US mobile daily active users come in at 642,000 for Muse against 231,000 for ChatGPT at the same point after its launch, and narrowed to iOS alone Muse still reads higher at 359,000, according to Apptopia.

The figures come from outside the company. Apptopia works from its own third-party estimates of downloads and active users, and TechCrunch writes that Meta has been asked for comment about Muse's early adoption.

The rankings have moved as well. Muse climbed from No. 2 overall on the US App Store right after launch to No. 1, Business Insider reported on Friday, which put it above ChatGPT, TechCrunch notes. Another firm, Appfigures, said at that point that Muse had crossed 1 million downloads.

Apptopia also published the audience overlap. More than 95% of Muse's users are Facebook users and 63% are Instagram users, the firm says. Meta ran the same playbook with Instagram's Threads, which now has more than 500 million users.

NVIDIA launches DSX Ready to qualify power and cooling products for AI factories

NVIDIA opened NVIDIA DSX Ready, a qualification program for the power and cooling products that go into AI factories, on September 21.

The yardstick is the DSX reference design. DSX Ready reviews whether a partner's product or solution meets the applicable NVIDIA DSX AI factory reference design requirements, and NVIDIA writes that category-specific requirements and review help builders evaluate offerings with greater confidence, reduce integration risk and move toward deployment.

Two categories open the program. It starts with battery energy storage systems (BESS) and cooling distribution units (CDUs), with Hitachi Energy, LG Energy Solution and Tesla listed as qualified BESS solutions and LG Electronics, LiquidStack and Vertiv as qualified CDU solutions. Additional categories across infrastructure and software will roll out over time.

The path differs by category. BESS providers run the required qualification tests and submit supporting data for NVIDIA review and approval within a defined qualification boundary. CDU providers use the CDU self-qualification suite to determine whether a specific offering meets the applicable NVIDIA functional requirements.

Qualification answers for the product. Passing it sits apart from site-level engineering and site-level stability, and NVIDIA writes that a qualified CDU may meet the relevant cooling criteria while the builder still evaluates how it will fit the planned facility.

The program is framed around physical constraints. As AI infrastructure expands, power, cooling, water, site and grid constraints are shaping what builders can deploy, and optimizing one part of an AI factory can shift the bottleneck elsewhere, NVIDIA writes.

Other Developments

Policy and Regulation

  • OpenAI published a position paper arguing that the United States should lead an international effort to develop global technical standards for frontier AI, including for recursive self-improvement (RSI). As AI systems take on more of the work of developing successive generations of AI, the company writes, they can increasingly drive RSI even while people remain involved, and OpenAI says fully autonomous RSI remains outside what is happening today and that whether and how to proceed must depend on preserving human control and on informed democratic choices about the benefits and risks. It names three problems international standards would address: fragmentation, where evaluations, reporting requirements and incident definitions conflict across nations; collective action, where each nation acting independently produces outcomes no nation wants; and uneven capacity, since frontier development and the expertise around it are distributed unevenly. The proposed mechanism leans on the network of AI safety institutes already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom, working through CAISI and national industry bodies, with common protocols for evaluating RSI-relevant progress, for human oversight of automated AI research, and for incident classification, tracking and reporting. The standards themselves would sit apart from licensing, mandatory prerelease review and approval requirements, with national governments deciding whether and how to write them into their own legal systems. Building standards for the next phase of AI (OpenAI)

Products

  • Google detailed the intelligence system built into its Googlebook laptop, centered on Gemini running on the device. Magic Pointer, summoned by a quick wiggle of the cursor, reads text, images and context on screen, so a marathon training plan highlighted on a web page can be mapped into Google Calendar, a suspicious email can be checked by its text and images, and a set of selected images can be worked on together in place. Rambler, the voice feature, takes a disorganized verbal brain dump and cleans up the stutters and filler, organizes action items into structured bullet points, and follows along when the speaker switches language mid-sentence. Create My Widget builds custom widgets from a spoken or typed description, and Google Antigravity ships with every unit alongside a full Linux terminal environment, so tools such as Claude Code or the Antigravity CLI run directly on the machine. Pricing starts at $899 with 12 months of Google AI Pro included, covering 5TB of cloud storage and the Gemini Advanced tools, pre-orders are open, and devices reach shelves on October 4 in the US and October 5 in Canada, the UK, Ireland, France, Germany and Australia. Googlebook's built-in intelligence reinvents the way you use your laptop (Google)

Litigation

  • Amazon.com Services LLC filed an amended complaint against Perplexity AI in the US District Court for the Northern District of California on September 21 (case 3:25-cv-09514-MMC, ECF No. 122, 41 pages). Amazon alleges that Perplexity built its AI agent Comet to operate inside the Amazon Store in precisely the ways Amazon's Conditions of Use forbid, sent the agent into password-protected areas after Amazon told Perplexity it was not permitted there, and shipped successive updates that defeated the technological measures Amazon deployed to stop it. On Comet for iOS, the filing alleges, Perplexity's own cloud servers took users' Amazon authentication cookies and browsed the Amazon Store directly, from the iOS launch on March 18, 2026 until at least May 11, 2026. The complaint pleads three counts — the Computer Fraud and Abuse Act (18 U.S.C. § 1030), California Penal Code § 502, and tortious interference with contractual relations under California law — and asks for preliminary and permanent injunctive relief barring AI-agent access to Amazon's protected systems, destruction of unlawfully obtained data, identification of every Amazon account accessed, monetary damages and attorneys' fees, with a demand for a jury trial. Amended Complaint #122 — Amazon.com Services LLC v. Perplexity AI, Inc. (CourtListener)

Security

  • NVIDIA published its framing of AI agent security as an engineering problem solved layer by layer. Models provide capabilities, harnesses organize context, tools and workflows, and runtime environments provide the infrastructure in which actions execute, and each part of that stack carries its own security responsibilities that call for controls at that layer. The worked example is an agent updating a customer record that meets malicious instructions in an attached document and attempts to export customer data to an unauthorized destination: a network policy should block the transfer, and protected logs should capture the attempted tool call, the authorization decision and the outcome so the security team can identify the tool used and the destination it reached for. Permission to update a customer record stays separate from permission to export it, and an agent can request additional access while the authority to grant it sits outside the agent. NVIDIA points to OpenShell, its open-source secure runtime that enforces policies outside the agent's reach and provides sandboxed execution, and writes that Cisco's DefenseClaw adds a governance layer on top of it while JFrog integrates with it to scan and verify agent skills. AI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent Stack (NVIDIA Blog)

Source: selected by the editors from the AI news inbox collected on September 22, 2026 (29 items, 23 primary and 6 secondary).