A proof can be machine-checked and still not be socially settled. OpenAI says an internal system using roughly 10,000 concurrent agents solved the Navier–Stokes existence and smoothness problem, then formalized the result in Lean. That is the defining story in AI news today—and the answer to the obvious question is: no, the mathematics community has not yet established that the 90-year-old problem is closed.

The claim arrives with unusually concrete artifacts and unusually awkward questions about concurrent research. Meanwhile, Meta has put an agent in a position to send email and spend money; Amazon has tied major purchase milestones to a Qualcomm warrant; Mistral has raised a record-sized European round; Washington is funding quantum manufacturing; and Arm is stretching one AI platform from phones to robots and data centers. The common thread is verification: when software can prove, purchase or move, you need to know what was checked and by whom.

AI News Today: OpenAI Says Its Agents Solved Navier–Stokes

OpenAI published a paper and a Lean formalization on September 8 claiming that a smooth, initially stationary three-dimensional fluid can develop a finite-time singularity under a smooth force. The company says the system reached the result after about 88 hours, with GPT‑6 Astra taking another 17 hours to formalize and verify it. The Navier–Stokes effort alone reportedly consumed 2.7 million agent messages and roughly 130 billion output tokens.

Those are company claims, not a verdict from mathematics. Lean can check whether a formal proof follows from encoded assumptions; it does not automatically prove that the encoding matches the Clay problem or that every imported lemma is appropriate. Independent experts must inspect both the argument and formalization. The Guardian reports that NYU mathematician Tristan Buckmaster questioned whether unpublished work he developed with Anthropic researcher Levent Alpöge may have influenced the system. OpenAI says neither agents nor researchers saw that work, while acknowledging it cannot completely rule out indirect model improvement from de-identified product data. For builders, the real milestone will be reproducible review—not the speed of the first announcement.

Meta’s Muse Turns the Personal Agent Into a Buyer

Meta introduced Muse as a personal agent that can use a browser, fill forms, negotiate, send email, book travel and keep working after the app closes. It is rolling out in the United States on iOS, Android and the web, with a free tier and paid plans for heavier use. Payments run through Stripe’s Link, which can generate a one-time-use card; Shop Pay and 1Password support are promised later.

Meta says every Muse runs in a dedicated cloud VM, a separate Sentinel agent approves outbound actions, and users confirm sensitive steps such as purchases. Treat those as design claims awaiting field evidence. The product’s importance is the permission boundary: once an assistant can act across inboxes, calendars and payment rails, a wrong answer becomes a transaction. Teams evaluating agents should demand scoped credentials, approval gates, audit trails and a clean way to revoke access. Our EU AI Act compliance guide is a useful companion for turning disclosure requirements into operational controls.

Amazon Gives Qualcomm a $60 Billion Path Into AI Infrastructure

Qualcomm and Amazon announced a multi-generation collaboration on customized AWS silicon for AI inference and optical links reaching 1.6 terabits per second. The more revealing detail sits in Qualcomm’s SEC filing: Amazon received a warrant for as many as 25 million Qualcomm shares at $161.26 each. Vesting is tied to commercial arrangements, orders and purchases, up to $60 billion in payments; 3.75 million shares vested at issuance based on initial commitments.

The ceiling is not a purchase guarantee. It is an incentive map showing how seriously both companies want an alternative path for inference compute and data-center networking. Qualcomm gains a hyperscale design partner beyond smartphones; AWS gains leverage over its silicon supply chain. The next evidence to watch is production timing, workload performance and actual purchase disclosures. For context on the capital intensity behind this contest, see our brief on Nvidia’s $96.2 billion quarter.

Mistral Raises €3 Billion to Keep a European AI Stack in the Race

Mistral says it raised €3 billion in a Series D led by Samsung Electronics, at a post-money valuation above €21 billion. Co-leads include EQT-managed Scaleup Europe Fund and existing investor PSG Equity. Mistral calls it the largest equity round completed by a European technology company; that superlative is the company’s characterization, while the round size and named investors are confirmed in its announcement.

The money is earmarked for frontier research, training compute, infrastructure and international growth. Sovereign AI is still a balance-sheet challenge: open-weight models require expensive chips, power and support. Mistral now has room to build, but funding is an input. Watch whether it converts strategic backing into competitive models, dependable enterprise deployments and economics that do not depend on Europe paying a sovereignty premium forever.

Washington Funds the Manufacturing Layer of Quantum Computing

The U.S. Commerce Department’s CHIPS for America program listed finalized awards of up to $100 million each for Rigetti, D-Wave and PsiQuantum on September 8, alongside a separate $375 million GlobalFoundries award. Quantinuum also announced a finalized $100 million award supporting trapped-ion hardware and U.S. supply-chain work with GlobalFoundries and Monarch Quantum.

The consequential shift is from laboratory promise to manufacturing repeatability. Quantinuum plans to use 300-millimeter wafer processes for ion traps and electronics, while Monarch works on scalable lasers and photonic components. None of the awards proves commercial-scale fault-tolerant quantum computing is imminent. They do show where government thinks the bottleneck sits: fabrication, packaging and components that can be produced consistently—not another heroic bench-top experiment.

Arm Connects Mobile, Robotics and Cloud AI on One Architecture

At Arm Everywhere China, Arm unveiled a platform-wide AI update spanning CSS for Mobile 2, a C2 Ultra CPU with SME2 matrix acceleration, a Mali G2-Ultra NX GPU with neural accelerators, the configurable Neoverse CSS N4 data-center platform and an AI Portal for developers. It also expanded Arm Total Design into physical AI with more than 80 participants and proposed a Robotics Capability Framework.

This is a portfolio announcement, so performance claims still need independent benchmarks and shipping partner products. Strategically, Arm is arguing that agents should inherit one software and architecture base as they move from cloud inference to phones, vehicles and robots. That could reduce deployment friction for developers, but it also raises the value of common capability labels and safety tests. Watch which chipmakers adopt CSS for Mobile 2 and N4—and whether the robotics framework becomes an industry vocabulary or a conference slide with excellent posture.

Watch & Learn

Editor’s note: The University of Bologna mathematics department’s “Getting Started with Lean — First Tactics and Examples” is a hands-on introduction to writing small machine-checkable proofs. It is best for technically curious readers who know basic logic; set aside about 90 minutes and keep Lean open if you want to follow the exercises rather than merely admire the syntax.

AI, Translated: Formal Verification

Formal verification means expressing a claim and its rules in a precise language, then using software to check that every logical step follows. For example, a Lean file can verify that a proposed mathematical proof reaches its stated theorem without an unsupported leap. It does not guarantee that the theorem statement faithfully represents the real-world question. You should care because agent outputs increasingly need evidence stronger than “the model sounded confident”—especially in mathematics, chips, security and automated transactions.

Try This Today

Goal: turn two source documents into a claim-evidence ledger in ChatGPT in about 10 minutes.

1. Upload the two PDFs or documents you want to compare. OpenAI says file uploads work on free and paid plans, subject to plan limits and workspace settings.
2. Ask for claims, exact source locations, contradictions and missing evidence—not a summary.
3. Open the cited pages yourself and mark each row verified, disputed or unresolved.

Copy-ready prompt: “Build a table with columns: claim, source file, page or section, exact supporting passage, status (confirmed/disputed/unresolved), counterevidence, and next verification step. Do not merge similar claims. If a page number is unavailable, say so. Separate what the source states from your inference, and finish with the three highest-risk unsupported claims.”

One Thing to Remember

The most important AI capability this week may not be generating an answer. It may be producing an audit trail strong enough for other people—and other machines—to challenge it.