OpenAI and the U.S. General Services Administration (GSA) announced a partnership to expand AI access for eligible federal, state, local, and tribal governments. The program offers $0 license fees, 50% off usage, and expanded cyber defense support.
DeepSeek officially launched V4.1 Flash on Thursday morning across web, App, and API, featuring a 552B-parameter MoE architecture with native multimodal support, which the company claims surpasses V4 Pro in performance, speed, and efficiency. The V4 Pro API will be retired on September 14, with requests automatically routed to V4.1 Flash, and DeepSeek Harness v0.1.5 is updated in sync. This move could further reduce inference costs and challenge top closed-source models on some benchmarks.
On September 10, 2026, Amap officially released ABot-Earth 0.7, a next-generation digital Earth model billed as the world's first 3D-native city world model. Trained on 3DGS data, it can generate kilometer-scale 3D city scenes in 10 minutes on a consumer-grade GPU, at one percent of traditional cost and about 1,000 times the efficiency. Amap aims to redefine digital Earth 21 years after Google Earth, with 3D maps covering over 190 countries and regions and interactive digital twin capabilities for physical AI such as robots and autonomous driving.
On September 10, Wang Mengdi, director of Princeton University's Center for AI Innovation, delivered a keynote at the 2026 Inclusion·Bund Summit main forum, arguing that while large models have absorbed vast human knowledge, they remain clearly limited in making original discoveries in basic science because they excel at finding the most likely answer in a probability distribution, whereas true discoveries often lie in the long tail. In collaboration with Princeton sociology professor Xie Yu, her team used large models such as ChatGPT and Claude for "simulated life" experiments, finding that models overestimate the most common outcomes and underestimate rare ones. Wang argues that the key to AI-driven autonomous scientific discovery lies in making the real world more verifiable, and her team is building an automated experiment platform and LabOS, an intelligent operating system for research laboratories.
Massachusetts Governor Maura Healey issued an executive order requiring data centers larger than 25 megawatts to supply their own clean power or contribute to a ratepayer protection fund. The mandate reflects growing state-level resistance to AI-driven data center expansion over energy costs and grid strain, following similar restrictions in Texas and New York. The AI industry, through the super PAC Leading the Future, is pushing back politically ahead of the midterm elections.
Harvey has raised $550 million at a $15.5 billion valuation in a round co-led by Diffusion and Lightspeed Venture Partners. The funding follows the launch of its first post-trained open-weight model and Harvey LAB, a Legal Agent Benchmark, and will support its efforts to help law firms, in-house legal teams, and professional services firms build and control their own AI capabilities at scale. The company said 80% of Am Law 100 law firms currently use Harvey, alongside numerous leading in-house legal departments, including five of the Fortune 10.
Mei Tao's team has officially launched Vivago R1, an AI video creation agent that underwent two months of beta testing. Positioned around "deterministic delivery," it aims to move AI video creation beyond the random generation mode. Vivago R1 uses a multi-agent architecture and supports conversational creation, capable of stably producing coherent videos of around 5 minutes while maintaining character and scene consistency. Previously, vivago.ai accumulated over 70 million global users, and this launch may push AI video creation toward a more controllable and accessible direction.