Global AI Weekly
Issue number: 107 | Tuesday, July 22, 2025
Highlights

Google acquires Windsurf licence after OpenAI deal fails
Google has secured top talent and the licensing rights from AI coding startup Windsurf after the startup's deal with OpenAI fell through. This move highlights Google's ongoing focus on bolstering its AI capabilities amid increasing competition in the field. The acquisition is seen as a significant step in strengthening Google's position in the AI landscape.
businesstimes.com.sg
Microsoft's new 'flash' reasoning AI model ships with a hybrid architecture, making its responses 10x faster with a '2 to 3 times average reduction in latency'
Microsoft introduced Phi-mini-flash-reasoning, a compact AI model in its Phi family that delivers responses 10 times faster. Built on an innovative hybrid architecture called SambaY, the model achieves a significant 2 to 3 times average reduction in latency. This advancement targets quicker and more efficient AI-powered interactions.
windowscentral.com
Amazon's $8 billion Anthropic investment rumors suggest it would rather sell AI infrastructure than compete with ChatGPT and Gemini
Amazon’s reported $8 billion investment in Anthropic highlights its strategy to focus on providing AI infrastructure instead of directly competing with AI platforms like ChatGPT and Gemini. With AWS already a key player in cloud computing, Amazon appears to be positioning itself as the backbone for AI innovations, offering the tools and infrastructure companies need to build and scale their own AI solutions. This approach aligns with AWS’s broader business model, prioritizing partnerships and empowering others in the AI ecosystem.
tomshardware.comResearch

Coding Triangle: How Does Large Language Model Understand Code?
This page explores how large language models comprehend and process programming code. It looks at the techniques and patterns these models use to understand syntax, logic, and structure in various coding languages. By examining these aspects, it provides insights into the capabilities and limitations of AI in handling complex coding tasks.
huggingface.co
One Token to Fool LLM-as-a-Judge
This paper explores the potential vulnerabilities of using large language models (LLMs) as judges in decision-making or evaluation tasks. It demonstrates how a single token can be strategically used to manipulate the outputs of these AI systems, raising concerns about their reliability and robustness. The discussion highlights the implications of such vulnerabilities and the need for improved safeguards when deploying LLMs in critical roles.
huggingface.co
A Practical Two-Stage Recipe for Mathematical LLMs: Maximizing Accuracy with SFT and Efficiency with Reinforcement Learning
This video or article explores a two-stage approach to improving mathematical large language models (LLMs). The first stage focuses on enhancing accuracy using supervised fine-tuning (SFT), while the second stage emphasizes efficiency through reinforcement learning techniques. It provides practical insights into the methods and their potential for optimizing mathematical reasoning in AI models.
huggingface.coVideo

Simon Hudson - Humans in AI
Simon Hudson, a retired tech enthusiast from Yorkshire, shares his passion for AI and its transformative impact on society in this Humans in AI episode. He discusses how artificial intelligence is not only advancing technology but also sparking important ethical conversations and encouraging personal growth. Simon also highlights the evolving dynamics of remote work and its role in shaping a new societal landscape.
youtube.com
Store and query vectors at scale in S3
Amazon S3 Vectors introduces cloud object storage with built-in support for storing and querying vectors at scale, cutting costs for these operations by up to 90%. This solution is designed to handle large vector datasets with durability and sub-second query performance, making it perfect for scaling billions of vectors while optimizing costs. Ideal for semantic search and AI applications with infrequent query needs, S3 Vectors offers a seamless way to manage large data volumes efficiently.
youtube.com
Using agent mode to refactor and iterate on a project
GitHub Copilot shines as a development teammate in this walkthrough with Simona, where she transforms a C++ console app into a multi-file Windows application. Using natural language prompts, she showcases how Copilot simplifies refactoring, fixing errors, and adding features with seamless tool support and real-time guidance. The demo highlights how these capabilities make coding more efficient and intuitive for developers at any level.
youtube.comArticles

Google’s Gemini refuses to play Chess against the Atari 2600
Google's Gemini AI declined to compete in a game of chess against the Atari 2600, signaling a surprising retreat in a match of capabilities. This decision follows claims that tools like ChatGPT and Copilot have already been surpassed, showcasing a shift in strategy. Gemini’s move reflects a focus on bigger challenges rather than revisiting outdated systems.
theregister.com
Exhausted man defeats AI model in world coding championship
After a grueling 10-hour coding marathon, a determined programmer emerged victorious against OpenAI in the world coding championship. The intense competition tested both human ingenuity and the capabilities of advanced AI, showcasing the resilience and creativity that humans bring to the table. Despite the challenges, the winner proved that humanity’s edge remains intact—for now!
arstechnica.comCode

stabilityai/stable-diffusion-3-medium · Hugging Face
Stability AI is committed to pushing the boundaries of artificial intelligence by focusing on open-source and open-science initiatives. Their Stable Diffusion 3 Medium model is a continuation of this mission, providing accessible tools for creative and innovative AI applications. It empowers users to explore and contribute to the evolving landscape of AI development.
huggingface.co
deepspeedai/DeepSpeed: DeepSpeed is a deep learning optimization library that makes distributed training and inference easy, efficient, and effective.
DeepSpeed is a powerful deep learning optimization library designed to simplify distributed training and inference while maximizing efficiency and effectiveness. It provides advanced features to accelerate large-scale model training, optimize resource utilization, and streamline deployment, making it an essential tool for AI developers.
github.comPodcast

The TWIML AI Podcast
The TWIML AI Podcast, hosted by industry expert Sam Charrington, explores the transformative impact of machine learning and artificial intelligence on businesses and everyday life. Featuring insights from leading ML and AI researchers, data scientists, engineers, and tech-savvy leaders, the podcast covers a wide array of topics, including deep learning, neural networks, natural language processing, and data science. It's an engaging resource for anyone interested in staying informed about cutting-edge technologies shaping the future.
open.spotify.com