The week of September 8-10, 2026 may be remembered as the moment AI crossed from assistive tool to autonomous scientific researcher - and simultaneously became something you might trust with your email, calendar, and credit card.
AI & ML News
OpenAI Claims AI Solved the 90-Year-Old Navier-Stokes Millennium Prize Problem
OpenAI announced on September 8 that an internal AI system produced an analytical proof showing the Navier-Stokes equations for three-dimensional fluid motion can develop a singularity in finite time - one of the seven Clay Millennium Prize Problems, each carrying a $1 million reward. The system deployed nearly 10,000 coordinating agents running for 88 hours on a model described as "significantly more capable than GPT-6 Astra." The proof, formalized in Lean and open-sourced on GitHub, describes a vortex spiraling inward while keeping energy finite. The announcement sparked an immediate credit dispute with NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge, who had been pursuing related work. Independent verification is now underway.
Meta Launches Muse - A Personal AI Agent on Its Own Secure Computer
Meta rolled out Muse to U.S. adults on September 9 - a personal AI agent that runs inside a dedicated Secure Virtual Machine with its own browser. A separate Sentinel agent acts as a gatekeeper, preventing any action from reaching the internet without explicit approval. Muse connects to Google Workspace, Ticketmaster, OpenTable, Spotify, and Apple Health, with purchases via Stripe Link (one-time card numbers, purchase protections included). Pricing ranges from free to $100/month. The launch represents Meta's most ambitious consumer AI product, arriving less than two weeks after the company's $18 billion social-media settlement.
Apple Unveils the iPhone Duo - Its First Foldable - Starting at $1,999
At Apple's "Surprise and Shine" event, new CEO John Ternus unveiled the iPhone Duo, a landscape-style foldable with a 7.7-inch display, titanium frame, A20 Pro chip, and 31 hours of battery life. The Duo supports Apple Pencil Pro and runs iPad-style apps that adapt when the screen folds. The event also brought the iPhone 18 Pro with the first variable aperture camera on an iPhone, and the Apple Watch Series 12 with real-time conversation transcription.
Cognition AI Hits $48B Valuation on $2B Raise
Cognition, maker of the autonomous software engineer Devin, nearly doubled its valuation from $26 billion to $48 billion in four months. Annualized revenue climbed to nearly $900 million, with Cognition expecting $4-5 billion by year-end. The round was led by a16z, Accel, Founders Fund, General Catalyst, and Avenir.
Harvey Raises $550M at $15.6B - Legal AI Gets Serious
Legal AI startup Harvey raised $550 million at $15.6 billion, with ARR past $400 million and 3,000+ customers including 80% of the Am Law 100. The company also acquired Guardrails AI for agent security and plans to build its own models. At ~39x revenue, the round confirms legal work as one of the first white-collar markets where AI startups are posting real revenue.
Big GitHub Releases
openai/NavierStokesAndEuler - Lean Formalization of the Millennium Prize Proofs
OpenAI open-sourced the machine-checked proofs for both the Navier-Stokes and Euler singularity results in Lean 4, hitting 1,600 stars in 48 hours. Each proof includes Jupyter notebook explanations. This represents a new kind of scientific artifact: produced by AI agents, verified by a formal proof assistant, and published as a repository.
affaan-m/ECC - Agent Harness Optimization System
ECC hit 250K stars, bundling skills, instincts, memory, and security practices for Claude Code, Codex, OpenCode, and Cursor into a single optimization framework. The system encodes engineering discipline as agent-level configuration.
humanlayer/skills - Human-in-the-Loop Agent Skills
HumanLayer's skills framework requires explicit human sign-off before every action execution. At 2.8K stars and growing fast, it addresses the core trust question in agent design: how to give AI real-world authority while maintaining control.
Under-the-Radar Finds
lnkiai/m3e-canvas (5.6K stars) bridges design and code: sketch Material 3 screens in the browser, export as structured prompts for AI coding agents. BraveOPotato/FckSignups (3K stars) is a curated list of open-source, in-browser tools that require zero signups. WorldFlowAI/everything-claude-code (2.5K stars) bundles the entire Claude Code best-practices toolkit into one installable set. sandblaze-ucb/ExploitGym from UC Berkeley benchmarks AI agents on real software exploitation tasks.
Why it matters. This week compressed three distinct AI narratives into 72 hours: AI can now tackle problems at the frontier of mathematics (Navier-Stokes), operate as trusted personal assistants with real-world authority (Muse), and become first-class citizens in product ecosystems (Cognition at $48B, Harvey at $15.6B). The question is no longer whether AI agents can do these things, but whether humans and institutions are ready for the consequences.