Headlines

1. Global Coalition Signs AI Compute Governance Treaty at UNGA
Heads of state from 40 nations agreed to a shared framework for fair access to data-center capacity, emergency model-disclosure protocols, and a $10B AI safety fund, marking the first binding global accord on frontier compute.
Source: Reuters

2. OpenAI Unveils GPT-6 Preview, Claims Breakthrough in Agentic Reliability
At the inaugural “AGI Day” in San Francisco, OpenAI demonstrated GPT-6 with self-correction loops and long-horizon planning, reporting a 12x reduction in task-failure rate on internal benchmarks compared to GPT-5.
Source: TechCrunch

3. Anthropic and DeepMind Independently Release Physical “World Models” for Robotics
Both labs published work showing foundation models that can accurately simulate real-world physics and object dynamics, enabling a new class of dexterous robot training without physical trials.
Source: The Verge

Model Releases

4. Google DeepMind Debuts Gemini 3.1 with 10M-Token Context Window
The updated Gemini line supports continuous video understanding, persistent memory, and a new Mixture-of-Experts architecture that cuts inference costs by 40% while doubling tool-call accuracy.
Source: Google DeepMind Blog

5. Alibaba’s Qwen 3.5 72B Open-Source Model Tops Global Leaderboards
Qwen 3.5’s instruct variant now outperforms rival open-weight models on code generation, math reasoning, and multi-step agentic workflows, and has been downloaded over two million times in its first week.
Source: Hugging Face

Industry

6. NVIDIA’s B300 “Blackwell Ultra” GPUs Begin Full Production
Announced at the Silicon Valley AI Summit, the B300 packs 288GB of HBM4 memory and dedicated “digital-twin” cores, with hyperscaler shipments expected to begin in October 2026.
Source: Bloomberg

7. Microsoft and Mistral AI Close $5B “Agentic Cloud” Joint Venture
The partnership aims to deploy regulated, sector-specific AI agents for healthcare and financial services across Europe, with data residency and EU AI Act compliance as core design features.
Source: Financial Times

Research Highlights

8. “DeepAnatomy: Generalist Medical Image Analysis via Self-Supervised Transfer”
A Stanford–Google collaboration shows that a single 7B-parameter model can outperform specialist systems across 30 imaging modalities, including X-rays, MRI, and pathology slides, reducing diagnostic turnaround in pilot hospitals by 22%.
Source: Nature Machine Intelligence

Compiled from multiple media reports.