Global AI Weekly
Issue number: 125 | Tuesday, November 25, 2025
Highlights
Ignite 2025: Microsoft Agents Take Center Stage
At Microsoft Ignite 2025, Microsoft unveiled a major leap into the agent-driven era with the launch of Microsoft Agent 365 — a control plane built to orchestrate AI agents across apps, cloud, and edge. Alongside, the company rolled out new agentic capabilities for Azure Copilot, enhanced governance, and a unified data + AI platform. The message: AI agents aren’t just assistants anymore — they’re integral to the next-gen enterprise.
news.microsoft.com
Gemini 3 Unveiled: A New Era of Intelligence
Gemini 3 is here, introducing a new level of intelligence to bring your ideas to life. This advanced model is designed to understand and assist with creativity, problem-solving, and innovation. It offers powerful capabilities to turn your concepts into reality with ease and efficiency.
blog.google
Google Denies Gmail Data Used for AI Training Amid Backlash and Misinformation
Google is under scrutiny after claims surfaced that it uses Gmail data to train AI models, leading to public calls for an opt-out option. The company has denied altering its policies, emphasizing that user data is only utilized for features like spam filtering, not for training AI systems like Gemini. The controversy highlights widespread concerns over AI data privacy and stresses the importance of understanding privacy settings.
webpronews.comResearch
Apple Machine Learning Research at NeurIPS 2025
Apple researchers are contributing to the advancement of AI and machine learning through foundational research, sharing their work to support the global research community. Their efforts aim to drive progress and innovation in the field.
machinelearning.apple.com
Think-at-Hard: Selective Latent Iterations to Improve Reasoning Language Models
This paper introduces "Think-at-Hard," a method designed to enhance reasoning in language models. The approach focuses on selectively applying additional latent iterations to challenging parts of reasoning tasks, improving accuracy and performance. By targeting only the hardest sections, it optimizes the reasoning process without increasing computational costs across the board. This work showcases its effectiveness through various experiments on reasoning benchmarks.
arxiv.org
Too Good to be Bad: On the Failure of LLMs to Role-Play Villains
This paper explores the challenges faced by large language models (LLMs) when attempting to role-play as villains. Despite their capabilities, these models often struggle to convincingly adopt villainous personas, revealing limitations in their alignment and behavior modeling. The study highlights intriguing insights into the ethical constraints and design of LLMs when performing morally complex tasks.
arxiv.orgVideo
Silicon Minds, Human Hearts - Monalisa Whalin
In this episode of "Human Hearts Silicon Minds," Monalisa Whalin, Principal VM Manager and AI Product Leader at Microsoft Azure OpenAI, shares her inspiring career journey and insights on cutting-edge AI advancements. She discusses transformative technologies like GPT-4, DALL-E, and Whisper, along with the potential of autonomous AI agents to enhance productivity. Monalisa also highlights the importance of sustainability, model evaluation, and building trust in AI while offering a glimpse into her personal favorite AI tools and this year’s biggest surprises in the field.
youtube.com
Ethan Ding - TextQL on Analytics Agents
This content introduces Ethan Ding's insights on TextQL, a tool designed to enhance the capabilities of analytics agents. It highlights how TextQL can streamline text-based data analysis, making it more efficient and accessible for various applications. The discussion emphasizes its potential to transform how analytics agents handle and interpret data effectively.
youtube.comArticles
LLMs can be easily jailbroken using poetry
This content highlights how poetry is being used as a clever method to jailbreak large language models (LLMs). By crafting specific poetic prompts, individuals can bypass restrictions and elicit unintended outputs from these AI systems. The approach showcases the unexpected vulnerabilities in LLMs and opens discussions around improving their safeguards.
theregister.com
How Copilot helps build the GitHub platform
This content explores how the Copilot coding agent has enhanced the GitHub platform, making it more efficient and transformative for developers. It highlights how Copilot streamlines coding processes, increases productivity, and empowers users to build more robust software. The focus is on the practical impact Copilot has on improving the overall GitHub experience.
github.blog
I Measured Neural Network Training Every 5 Steps for 10,000 Iterations
This article explores the insights gained from measuring neural network training every 5 steps over 10,000 iterations. By closely observing the training process, it highlights how features form and evolve within neural networks, shedding light on the dynamics of their learning patterns. The detailed approach provides a clearer understanding of how networks develop their representations through iterative updates.
towardsdatascience.com
AI-Generated Code Creates New Wave of Technical Debt
A recent report by Ox Security, titled "Army of Juniors: AI-Generated Code Creates New Wave of Technical Debt," highlights a key concern with AI-generated code. While such code is highly functional, it often lacks the architectural judgment necessary for long-term software sustainability. This shortfall is contributing to technical debt, posing challenges for companies relying heavily on AI tools for coding.
infoq.comUpcoming Events
AgentCon 2025
After an exciting set of AgentCon events before the summer holidays, AgentCon is happening in more locations during the second half of the year. The AI Agents World Tour is a series of one-day conferences for developers focused on autonomous AI agents. Spanning cities like San Francisco and Singapore, these events gather top engineers, researchers, and creators shaping the future of AI technologies. It’s an opportunity to connect, learn, and innovate in the fast-evolving world of autonomous AI. Check all events and register for an event near you!
globalai.communityCode
deepseek-ocr
DeepSeek-OCR is a vision-language model designed for efficient optical character recognition (OCR). It focuses on token efficiency, making it effective in extracting text from images while using fewer computational resources. This model offers a streamlined approach to OCR tasks, enhancing accuracy and performance.
ollama.com
theNetworkChuck/docker-mcp-tutorial: Complete tutorial materials for building MCP servers with Docker
This repository provides comprehensive tutorial materials for creating MCP servers using Docker, based on NetworkChuck's video guide. It covers everything you need to set up and manage your own server environment efficiently. The materials include detailed instructions, scripts, and examples to help you get started with Docker and MCP servers.
github.com
Using your LLM provider API keys with Copilot - GitHub Docs
Discover how to enhance GitHub Copilot by integrating it with your preferred custom models using your own LLM API keys. This guide walks you through the process of configuring and making these models accessible for organizations within your enterprise account. Simplify collaboration and optimize Copilot to better suit your team's specific needs.
docs.github.comPodcast
Generative Now | AI Builders on Creating the Future
Generative Now is a weekly series by Lightspeed that showcases the inspiring stories and strategies of today’s leading AI companies and innovators. Hosted by Lightspeed Partner Michael Mignano, the podcast features conversations with founders, engineers, designers, and other key players using AI to revolutionize their industries. Tune in to explore how these trailblazers are shaping the future of work and leveraging AI to elevate their products.
open.spotify.com