OpenAI Says Internal AI System Resolved the Navier–Stokes Problem
An internal AI system produced a proof that the Navier-Stokes equations can develop a singularity in finite time, formalized in the Lean proof assistant. OpenAI announced this result on September 8, 2026, claiming it resolves one of the Clay Mathematics Institute's Millennium Prize Problems.
NSA, CISA, FBI Warn China-Based AI Firms Distill US Frontier Models
China-based AI firms have been systematically extracting capabilities from U.S. frontier models through knowledge distillation campaigns since late 2024, according to a joint advisory released September 8, 2026 by the NSA, CISA, and FBI. The advisory, designated AA26-251A, warns that these extraction efforts operate at industrial scale.
Meta debuts its Muse AI agent. Will consumers trust it?
Meta released a personal AI agent called Muse that requests access to users' email, calendars, payments, and health services. The agent represents a significant expansion of Meta's consumer AI capabilities and a test of user trust regarding data sharing with the company.
Cognition Raises Over $2B Series E at $48B Valuation to Scale Devin Agents
Cognition raised over $2 billion in Series E funding at a $48 billion valuation to expand its AI software-engineering operations. Andreessen Horowitz and Accel led the round alongside existing investors Founders Fund, General Catalyst, and Avenir, with additional participants including Benchmark and Bessemer.
Arm launches Total Design for Physical AI and robotics framework
Arm introduced Total Design for Physical AI and a robotics framework to standardize automated systems across industries like mining, agriculture, and manufacturing. Physical industries represent an estimated $200 billion annual compute opportunity by the 2030s.
This AI entrepreneur is developing agents that can plan ahead for the unexpected
Danijar Hafner's stealth-mode startup is developing AI agents capable of planning ahead for unexpected situations. The San Francisco company currently operates with minimal staff and furnishings in its SoMa district office.
Coca-Cola uses AI to improve retailer ordering in Malaysia
An AI-powered recommendation system analyzes previous orders, seasonality, weather, and purchasing patterns to suggest products and quantities for Malaysian retailers. The Perfect Basket feature on the Coke Buddy platform currently supports approximately 39,000 retail outlets.
AI weather forecasting enters the energy market as Google targets grid operators with WeatherNext 3
An AI weather forecasting model predicts wind speed at 100 meters above ground, cloud cover, and surface sunlight with hourly updates. Energy traders, grid operators, and renewable developers currently purchase similar forecasting services from other companies.
Reducto Releases r-1: A Single Pass Document Parsing Model That Cuts Errors 20% at 1 Cent Per Page
A single-pass document parsing model called r-1 reduces errors by 20% while processing full pages through combined OCR, layout detection, tables, and formatting in one operation. The model costs 1 cent per page compared to the legacy system's 3 to 6 cents per page.
XPENG Commissions Humanoid Robot Lines as IRON Walks Off Production
XPeng has commissioned production lines for humanoid robots and completed the first unit of its IRON robot, which walked off the production line autonomously. The facility represents a transition from research-and-development prototyping to line manufacturing for the next-generation IRON robot model.
Firmus Signs OpenAI as Anchor Customer for Malaysian AI Factories
OpenAI will contract dedicated AI compute capacity from two Firmus AI Factory sites in Malaysia under a multi-year partnership announced on September 8, 2026. This agreement brings Firmus Technologies' total contracted capacity across all customers to beyond 900 MW.
Alibaba Cloud, Cambricon Join PyTorch Foundation, Ant Group Takes Gold Seat
Alibaba Cloud and Cambricon joined the PyTorch Foundation as Platinum members while Ant Group joined as a Gold member on September 8, 2026. The announcement was made at KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 in Shanghai.
OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device
A 2.52 billion parameter language model called MiniCPM5-2B was released with a 131,072 token context window and scores averaging 53.9 across 34 benchmarks. The model uses 400 billion tokens of deep-thinking training paired with reinforcement learning from 16 merged expert models and runs on standard architectures like llama.cpp and Ollama without requiring custom code.
Axis Robotics Releases AXIS: A Browser-Based Data Engine With 207 Robot Manipulation Tasks and 50,129 Trajectories
Axis Robotics released AXIS, a browser-based data engine that collects robot manipulation demonstrations on web interfaces while processing computations on backend GPUs. The system generated 207 tasks with 50,129 verified Franka trajectories, improving π0.5 performance from 83.9 to 88.8 on LIBERO-Plus.
Grupo Financiero Inbursa Adopts Harvey Across Its Legal Organization
Grupo Financiero Inbursa deployed Harvey's legal AI platform across its litigation operations, with plans to expand to additional legal workflows. The Mexico City-based financial institution is using the platform to modernize legal operations and enable lawyers to work more efficiently on complex legal matters.
9kV Solid-State Transformer Anchors Hyosung’s U.S. AI Grid Push
A 22.9kV solid-state transformer developed by Hyosung Heavy Industries in 2022 is being deployed to power U.S. AI data centers. The company is combining this technology with completed American production facilities for ultra-high-voltage transformers and circuit breakers.
MG Ship adds AI route optimisation as logistics returns accelerate
MG Ship introduced an AI route optimization module that pairs automated routing algorithms with carrier recommendation systems for international trade corridors. The technical system targets global retailers and commercial shippers seeking improved logistics deployment efficiency.
The Download: the hunt for underground hydrogen and more rogue OpenAI agents
Exploration efforts are searching for underground stores of hydrogen gas to provide zero-carbon fuel sources. Companies and researchers are conducting a flurry of investigations to locate these natural hydrogen deposits beneath the Earth's surface.
Google, Cathay Pacific Expand Contrail Avoidance Trials in Asia-Pacific
Google and Cathay Pacific expanded AI-powered contrail avoidance trials on ultra-long-haul flights across Asia-Pacific, reducing contrail warming impact by an estimated 40%. Cathay Pacific became Google's first commercial airline partner in Asia testing contrail avoidance on routes exceeding 16 hours and spanning 13,000 kilometers.
IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B
IFM released K2 Horizon, a suite of six Apache 2.0 models ranging from 0.9 billion to 375 billion parameters. The models include variants sized at 0.9B, 3.7B, 7B, 32B, 36B, and 375B parameters.
H Company Releases NeoMME: A Family of 260M and 800M Single-Tower Multimodal Encoders That Drop the Vision Tower and Causal Decoder
H Company released NeoMME, a family of multimodal encoders with 260M and 800M parameters that process text and 32×32 image patches in a single Transformer without separate vision towers. The 260M model achieved 0.523 nDCG@10 on ViDoRe v3 and indexes at 51.3 pages per second on one L40S GPU.
In “An Alien Mind,” OpenAI’s Jakub Pachocki Urges Shared Safety Bars
OpenAI Chief Scientist Jakub Pachocki published an essay on September 6, 2026, arguing that no AI lab has adequately solved alignment and monitoring to scale responsibly at maximum speed. He called for voluntary slowdowns and international coordination to establish shared safety standards before advancing further AI development.
Meta FAIR Introduces AI Research Preference Models (RPMs): Ranking ML Experiments Before Spending GPU Hours
Research Preference Models rank unexecuted ML experiments before execution, allowing selection of the most promising candidate from 15 proposals. On AIRS-Bench, this approach improved average normalized scores from 0.684 to 0.729 and achieved baseline results in approximately 15 hours instead of 24 hours.
Amazon Sets Final Closure of Mechanical Turk, Its Crowdsourcing Marketplace
Amazon's Mechanical Turk crowdsourcing marketplace will permanently close on September 30, 2026. The platform previously connected businesses with distributed workers for small digital tasks that Amazon called Artificial Artificial Intelligence.
OpenAI Hits Goal of Building an ‘Automated Research Intern’
OpenAI developed an automated research system that achieves 3.1 agent-workdays of effort for every workday of human labor. The company says it reached this goal by September 2026, as previously announced in fall 2025.
What Is Model Routing? How AI Systems Choose the Right Model for Every Request
Model routing systems automatically select among different models or configurations based on request-specific factors like capability, cost, and latency. These systems balance trade-offs between performance quality, response speed, availability, and operational expense for each individual query.
UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents
UC Berkeley researchers released CUA-Lite, an open platform that unifies computer-use agent training by standardizing agents, environments, traces, and evaluation frameworks into one action space and data schema. The platform reduces virtual machine size from 4.1 GB per task to a single Docker container at 0.9 GB.
Perplexity Details Its GPU Embedding Stack: How Ivy, Tulip and ROSE Serve pplx-embed
Perplexity published details about its GPU embedding infrastructure, including systems named Ivy, Tulip, and ROSE that serve its pplx-embed model. The infrastructure focuses on optimizing how embedding models run across indexes at scale to improve retrieval quality in AI search.
GitHub Introduces Project HydraFusion: Runtime Multi-Model Orchestration That Builds a Workflow Per Coding Task in Copilot CLI
GitHub's Project HydraFusion automatically selects between three execution patterns,Single, Cascade with quality gates, and Critique with cross-family reviewers,to optimize workflow selection for each coding task in Copilot CLI. The system treats workflow routing as an optimization problem rather than relying on a single model picker.
Nous Research Adds One-Click Local Model Setup to Hermes Desktop
Hermes Desktop now offers one-click local model setup that automatically detects hardware, selects compatible models, and configures llama.cpp with a 4-bit quantization floor. The application matches available GPU resources against its catalog to download and install the highest-quality model build supporting a minimum 64K context window.