Global AI Weekly
Issue number: 133 | Tuesday, January 20, 2026
Highlights
Apple will pay billions for Gemini after OpenAI declined
A recent report highlights Apple’s partnership with Google to integrate Gemini, a cutting-edge AI model, into the new Siri, following OpenAI’s decision to decline. This move suggests Apple is making significant investments, amounting to billions, to enhance Siri's capabilities and compete in the AI space. The collaboration aims to deliver a more advanced and seamless user experience.
9to5mac.com
gRPC as a custom transport for MCP
Google Cloud is collaborating with MCP maintainers to enhance the MCP SDK by introducing support for pluggable transports. This update includes enabling gRPC as a transport option without the need for transcoding. These advancements aim to provide developers with more flexibility and efficiency in their applications.
cloud.google.comResearch
Reimagining LLM Memory: Using Context as Training Data Unlocks Models That Learn at Test Time
The NVIDIA Technical Blog explores a new approach to enhancing large language models (LLMs) by using context as training data. This innovative method allows models to learn during test time, enabling them to adapt and improve performance with extended memory capabilities. By leveraging this concept, LLMs can process larger volumes of information and maintain more detailed conversation contexts.
developer.nvidia.com
Learnable Multipliers: Freeing the Scale of Language Model Matrix Layers
This paper introduces the concept of learnable multipliers, a method designed to optimize the scale of matrix layers in language models. By allowing the scale to adjust dynamically, the approach aims to improve efficiency and performance in these models. The discussion sheds light on how this technique could enhance large-scale language model training and deployment.
huggingface.co
Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models
The paper introduces Nemotron-Cascade, a scalable approach to cascaded reinforcement learning designed for general-purpose reasoning models. It focuses on enhancing the performance and flexibility of such models by employing a cascade framework that optimizes decision-making through multiple stages. The method aims to improve efficiency and adaptability in tackling complex reasoning tasks across various scenarios.
arxiv.orgVideo
Human Hearts, Silicon Minds - Scott Hanselman
In this episode of Silicon Minds Human Hearts, Scott Hanselman from Microsoft shares his experiences from 30 years of working with technology and his insights on the evolving relationship between humans and AI. He discusses the challenges of balancing rapid technological advancements with digital ethics and explores the productivity paradox introduced by AI. Scott also highlights compelling AI applications in personal life and education while offering practical advice on staying informed in a fast-moving AI landscape.
youtube.comArticles
How Nano Banana got its name
We explore the quirky backstory behind the name "Nano Banana," one of Google DeepMind's well-known models. From a fun brainstorming process to the model's unique connection to its name, discover how this playful moniker came to be. It’s a lighthearted look into the creative side of tech innovation.
blog.google
Announcing Agent Academy: Operative
This update introduces Agent Academy: Operative, the next stage in the learning path for building agents in Copilot Studio. Following the success of the Recruit curriculum, this new step offers more advanced resources to help makers and developers enhance their skills. The goal is to continue supporting the growing community of learners with tools and knowledge for creating powerful AI solutions.
devblogs.microsoft.com
Choosing the Right Multi-Agent Architecture
This post discusses the key moments when multi-agent architectures are needed and highlights four common patterns observed in their use. It explains how LangChain provides effective tools and solutions for building and managing multi-agent systems. The content serves as a guide to help you choose the right architecture for your specific needs.
blog.langchain.com
OpenAI Codex with Ollama
OpenAI's Codex can now be utilized with Ollama's CLI, enabling seamless interaction with your working directory. This integration allows Codex to read, modify, and execute code using open-weight models like gpt-oss:20b and gpt-oss:120b, among others. It offers a versatile way to enhance coding workflows with advanced AI capabilities.
ollama.com
GitHub Copilot now supports OpenCode
GitHub Copilot has introduced full support for authentication with OpenCode through a formal partnership. This means users with Copilot Pro, Pro+, Business, or Enterprise subscriptions can now seamlessly integrate and use their subscriptions with OpenCode. The collaboration aims to enhance accessibility and usability for developers working across platforms.
github.blogUpcoming Events
AgentCon - The AI Agents World Tour Continues in 2026
AgentCon continues into 2026 with the AI Agents World Tour—one-day, developer-focused conferences dedicated to autonomous AI agents. Building on a successful run of events, the tour expands to even more cities worldwide, from San Francisco to Singapore and beyond. Join leading engineers, researchers, and builders to explore cutting-edge agent architectures, real-world use cases, and emerging best practices. Connect with the global AI community and help shape the future of autonomous AI.
globalai.communityCode
ronantakizawa/a11ymcp: MCP Server for Web Accessibility Testing
This project provides an MCP Server designed for web accessibility testing, showcasing its utility with over 5000 downloads and ranking as #20 on ProductHunt. It aims to help developers ensure web accessibility in their projects effectively. Perfect for those looking to improve inclusive design in their web applications.
github.com
naklecha/simple-llm: ~950 line, minimal, extensible LLM inference engine built from scratch.
A lightweight and minimal LLM inference engine created entirely from scratch in about 950 lines of code. Designed to be extensible, this project provides a straightforward foundation for working with large language models. Ideal for those interested in understanding and building custom inference engines.
github.comPodcast
This Day in AI Podcast
This Day in AI Podcast features Michael and Chris Sharkey, two everyday tech enthusiasts navigating the world of artificial intelligence with humor and relatability. The podcast offers an hour of casual conversations about AI experiments, amusing mishaps, and average yet entertaining advice on AI tools. With no expert credentials but plenty of curiosity, the Sharkey brothers share stories, prank calls, and even AI-created songs, making AI approachable and fun for everyone. New episodes come out whenever inspiration strikes!
open.spotify.com