Global AI Weekly - Issue 163 - Global AI Community

In this issue

Highlights

2 stories

Microsoft Wants to Break Nvidia’s Grip With 300,000 AI Chips

Microsoft is preparing a major production push for its next-generation Maia 300 AI chip, potentially unveiling it as soon as September. The company is reportedly discussing capacity with TSMC for more than 300,000 chips in 2027, with ambitions to scale even further. The move could reduce Microsoft’s reliance on Nvidia, lower the cost of running AI workloads in Azure, and put its custom silicon strategy closer to those of Google and Amazon.

macrumors.com

Apple Is Building a China-Only AI Model

Apple has trained its own large language model specifically for China, with help from Alibaba. The move marks a shift from Apple’s previous plan to rely mainly on local third-party models for Apple Intelligence. Alibaba’s Qwen will still play a role, but Apple now appears to want more control over the underlying AI experience. It is another sign that regulation is pushing global tech companies toward increasingly regional AI strategies

In this issue

Research

2 stories

arxiv.org

Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill

Spark-to-Paper takes a research idea and runs almost the entire scientific workflow inside a coding assistant. Its 13 composable skills search literature, plan and execute experiments, revise claims based on results, generate figures and assemble the final manuscript. Across eight research topics, the system reported 99.5% citation validity, while its integrity checks dramatically improved fabrication detection. It is an interesting glimpse at what automated AI research might actually look like.

arxiv.org

ToolHazard: Scaling Adversarial Environments for Security Evaluation and Alignment of LLM-based Agents

Give an AI agent tools and access to external data and attackers gain a much bigger playground. ToolHazard automatically creates adversarial environments designed to expose agents to indirect prompt injections hidden inside the systems they interact with. The resulting ToolHazard-Bench reveals substantial vulnerabilities, including differences depending on where and when malicious instructions appear. The researchers also show that training on these generated attacks can improve agent security without sacrificing normal task performance.

In this issue

Articles

2 stories

thenewstack.io

Qwen3.8-27B Wants to Put Opus-Class AI on Your Mac

libaba’s new open-weight Qwen3.8-27B model is pushing powerful local AI closer to everyday developer hardware. The 27B multimodal model targets coding, vision, research and agentic workloads, and can run locally with tools such as llama.cpp and LM Studio. Early tests show strong performance, but memory requirements, quantization, inference speed and context handling still matter. The result is an increasingly capable cloud alternative for developers who want more control, privacy and offline access.

openai.com

OpenAI Quietly Makes ChatGPT Smarter and Free to Use

OpenAI is updating GPT-5.6 Sol in ChatGPT with more focused answers, tighter formatting, and better factual reliability. Plus and Pro users also get a new control for choosing how much reasoning effort goes into a response, while Instant and Thinking become more consistent. Free users get an even bigger change: GPT-5.6 Luna becomes the default, with unlimited text chats and a new Think button for harder questions.

In this issue

Upcoming Events

2 stories

nvidia.com

October 20–22: NVIDIA GTC Berlin

Europe's biggest AI moment is happening this autumn. GTC Berlin brings together developers, researchers, and industry leaders to go deep on the full five-layer AI stack, from energy, chips, and infrastructure to open models and physical AI. The headline keynote by Jensen Huang takes place at the iconic Tempodrom on 21 October, with three days of sessions, hands-on labs, and networking at STATION-Berlin. For developers shipping agentic workflows or navigating open-source AI, GTC Berlin is where it all comes together. Eligible attendees get 25% off.

globalai.community

Upcoming Events Around the World

From local meetups to full-day conferences, the Global AI Community brings developers, architects, founders, and AI enthusiasts together to learn, build, and connect. Discover hands-on workshops, expert talks, community gatherings, and flagship events like AgentCon, AgentCamp, and Global AI Construct. Whether you are exploring AI for the first time or building production-ready agentic systems, you will find opportunities to sharpen your skills, meet like-minded people, and become part of a thriving global community.

In this issue

Code

2 stories

github.blog

Put GitHub Copilot Inside Your Java App

GitHub’s Copilot SDK is not just for IDEs. This tutorial shows Java developers how to create agent sessions directly from server-side applications, register tools, send prompts and consume structured responses. The walkthrough builds a complete Jakarta EE 11 application with concurrent agents running on virtual threads. Interestingly, the SDK also supports bringing your own model provider, including OpenAI, Azure and Anthropic, making it useful far beyond traditional Copilot integrations.

developers.googleblog.com

Run a Local AI Agent on a Raspberry Pi

Google shows how far edge AI has come by running Gemma models locally on a Raspberry Pi 5 using LiteRT. The tutorial covers model inference, CPU and GPU acceleration, object detection, speech recognition and an autonomous robotics pipeline using Reachy Mini. Developers can also add LiteRT skills to coding agents to automate conversion, quantization, benchmarking and inference workflows. The result is a practical path to capable AI applications with no cloud dependency.

In this issue

Podcast

1 story

open.spotify.com

What Happens When Your AI Coworker Remembers Everything?

Lindy founder Flo Crivello joins The Cognitive Revolution to explain Lindy Teammate, an AI coworker that lives inside Slack and continuously builds shared knowledge about a team. The discussion gets surprisingly technical, covering memory agents, file-based context, retrieval, sandboxing, evaluations and prompt optimization. Crivello also explains why Lindy currently uses DeepSeek, while simultaneously arguing that access to Chinese AI models raises difficult economic and policy questions.