Hey 👋

Imagine waking up one day and deciding to give away $200 billion, on purpose. That’s what Bill Gates just did. He’s pledged to donate nearly all his wealth by 2045, then shut down the Gates Foundation forever. No legacy empire. No billion-dollar endowment. Just one last moonshot to rewrite the future for global health, poverty, and education.

In a world where tech leaders are stockpiling power, Gates is choosing a full reset. No cliffhangers. Just closure.

Let’s see what the rest of the AI world was up to this week 👇

What would be the format? Every week, we will break the newsletter into the following sections:

• The Input - All about recent developments in AI

• The Tools - Interesting finds and launches

• The Algorithm - Resources for learning

• The Output - Our reflection

Table of Contents

• The Input

  • 📐Google’s AlphaEvolve Cracks Complex Math and Boo …

  • 🩺 OpenAI Boosts Health AI with HealthBench + Adds …

  • 🐰CodeRabbit Brings Free AI Code Reviews into Your …

  • 🛍️ Google Turns 2D Images into 3D Shopping with G …

  • 🧑‍💻 AG-UI Is Here: A Lightweight Protocol for Ag …

  • 🧠 Qwen Chat’s Deep Research Is Now Live

  • ⚡ MetaShuffling Makes Llama 4 Way Faster and Smart …

  • 🚀 Google Launches AI Futures Fund for Startups

• The Tools

• The Algorithm

• The Output

The Input

📐Google’s AlphaEvolve Cracks Complex Math and Boosts Efficiency

Google just introduced AlphaEvolve, an AI coding agent that uses Gemini and evolutionary search to create and improve algorithms for complex math and computing challenges.

What’s New:

• Code that evolves: AlphaEvolve mutates, scores, and refines programs over generations, automatically discovering novel solutions from scratch.
  

• Powered by Gemini: Combines Gemini Flash for wide exploration with Gemini Pro for deep analysis and refinement.
  

• Auto-evaluated: Solutions are checked for correctness and performance without human involvement.
  

• Real-world impact: Already improving Google’s data center efficiency, cutting compute waste by 0.7%, speeding up AI training kernels by 32.5%, and contributing chip design optimizations for next-gen TPUs.
  

• Math breakthroughs: Found new matrix multiplication algorithms and advanced 20% of 50+ open math problems, including the 11-dimensional kissing number.

AlphaEvolve is currently used internally at Google but an Early Access Program for select academic users is in the works. If you’re interested, you can register through Google’s official form. The team is also exploring ways to make it more broadly available in the future. (source)

🩺 OpenAI Boosts Health AI with HealthBench + Adds Fidji Simo as CEO of Applications

OpenAI just released HealthBench, a tough new test to see how well AI can handle real medical conversations. At the same time, they brought in Fidji Simo as CEO of Applications to help turn AI research into real-world products.

What’s New:

• Real Doctor Input: HealthBench was created with help from 262 doctors and includes over 5,000 medical conversations. It checks how accurately AI can answer questions in important healthcare situations.

• Two Special Sets of Tests: HealthBench includes two special groups of medical cases for testing AI -  one with questions where many doctors agree on the best answers (Consensus), and another with tougher questions that even current AI models find hard to answer correctly (Hard). This helps highlight where AI is doing well and where it still needs improvement.

• Top Performers: OpenAI’s newest models like GPT-4.1 and o3 are leading, sometimes even beating human experts.
  

• AI as a Judge: GPT-4.1 helps score AI answers, agreeing closely with real doctors. (source)

Leadership Update: Fidji Simo, former Instacart CEO and OpenAI board member, joins as CEO of Applications to lead scaling and responsible deployment, reporting directly to Sam Altman. (source)

🐰CodeRabbit Brings Free AI Code Reviews into Your IDE

CodeRabbit just made coding smoother with real-time AI-powered code reviews inside your favorite editor. It’s like having a senior engineer watching your back before you even push your changes.

What’s New:

• Stay in Flow: Get instant feedback on your uncommitted changes and fix bugs or refactor suggestions with just one click -  no need to leave your coding environment.
  

• Smart Context-Aware Reviews: The AI understands your code changes, dependencies, and coding style to catch over 95% of bugs and logic errors early.
  

• Multi-Language Support: Works across JavaScript, Python, Go, C++, Java, Rust, PHP, and more -  so you can rely on it regardless of your stack.
  

• AI Hallucination Detection: Flags potential AI-generated mistakes or weird logic before it becomes a problem in your code.

Unlike many tools that only review committed code or focus on static analysis, CodeRabbit reviews your in-progress, uncommitted changes live,  meaning it catches problems before they even hit your repo, speeding up your workflow and reducing bugs early. (source)

🛍️ Google Turns 2D Images into 3D Shopping with Generative AI

Google is using generative AI to transform flat product photos into immersive 3D shopping experiences- making online browsing feel a lot more like in-store discovery.

What’s New:

• From 2D to 360° in 3 Images: New AI models now generate high-quality 3D visualizations from as few as three product photos- cutting down cost and complexity for businesses.
  

• Built with Veo: The latest breakthrough leverages Veo, Google’s video generation model, to create smooth, 360° spins with realistic lighting, textures, and material effects- even from minimal input.
  

• 3 Generations of Progress:
  
  

  • NeRF (2022): Enabled 360° views from 5+ images using neural radiance fields.
    

  • View-Conditioned Diffusion (2023): Used DreamFusion-style training to scale to more complex shapes and fewer photos.
    

  • Veo (2024): Generalizes across categories- furniture, fashion, electronics- and handles reflections, lighting, and geometry with minimal data.
    

• Live on Google Shopping: These AI-generated 3D views are already powering product listings- especially for shoes, sandals, and apparel.
  

Google’s tech is helping bridge the gap between screen and shelf- making online shopping more tactile, personalized, and visually rich. (source)

🧑‍💻 AG-UI Is Here: A Lightweight Protocol for Agentic Apps

A new open protocol called AG-UI is here to standardize how AI agents connect with front-end apps. Lightweight and event-based, it’s designed to simplify building agentic user experiences- without locking you into a single stack.

What’s New:

• Plug-and-Play Simplicity: AG-UI defines 16 standard event types and structured inputs for easy frontend integration- compatible with WebSockets, SSE, webhooks, and more, thanks to a flexible middleware and reference HTTP connector.
  

• Framework-First & Developer-Tested: Designed with real-world input from CopilotKit devs and integrated with agent frameworks like LangGraph, Mastra, CrewAI, and AG2- reflecting actual developer workflows.
  

• Out-of-the-Box Ecosystem Support: Fully supported in top frameworks today, with contributions open for others like OpenAI SDK, AWS Bedrock, and Vercel AI SDK.
  

• Packed with Core Features: Real-time agentic chat, tool use, delta streams, structured UIs, human-in-the-loop collaboration, and more- baked in.
  

• Production-Ready Clients: React client via CopilotKit is available now, with messaging integrations (WhatsApp, WeChat, RCS) coming soon in collaboration with AWS.

AG-UI addresses the challenge of seamlessly connecting advanced AI agents with user-facing apps by providing a standardized, event-driven protocol that simplifies integration and enables real-time, interactive experiences. With 16 standard event types and support for features like streaming and state sync, it reduces reliance on custom solutions while ensuring broad compatibility across AI frameworks and platforms. (source)

🎮 Try it Out: AG-UI demo

🧠 Qwen Chat’s Deep Research Is Now Live

After weeks of testing, Qwen Chat’s Deep Research feature is now officially available for all users- joining the ranks of AI models like ChatGPT, Bard, and Bing AI that offer advanced deep research capabilities to deliver comprehensive, nuanced insights. While some variants of Qwen may have offered research-like functions before, this release marks a full rollout of their dedicated Deep Research mode.

What’s New:

• Ask Anything, Get Specific: Start with a broad question like “Tell me about robotics”. Qwen will help narrow it down- history, theory, or real-world use.
  

• Surprise Me Mode: Not sure what you want? Qwen can choose a direction for you and still deliver something insightful.
  

• Hands-Free Reports: While you sip your coffee ☕, Qwen builds a detailed answer, personalized and easy to digest.
  

• For Work or Play: Great for learning, researching, or just exploring random ideas with AI that feels like a helpful study buddy.

🔗 Try Deep Research on Qwen Chat

⚡ MetaShuffling Makes Llama 4 Way Faster and Smarter

Meta just launched MetaShuffling, a new way to make their big Llama 4 models work faster and smarter. These models use a mix of experts (called Mixture-of-Experts or MoE) to handle huge amounts of data -  but until now, their method of processing slowed things down.

What’s New:

• Better Token Sorting: Instead of wasting time and computer power padding or cutting data awkwardly, MetaShuffling groups similar tasks (“tokens”) together. This means the GPU can work more efficiently without waiting around.
  

• Smarter Work Splitting: It uses special code (GroupedGEMM kernels) to split the work evenly, save memory, and keep important info handy -  like tidying your workspace so you’re never searching for what you need.
  

• Made for Big Setups: Works smoothly on multiple GPUs and even multiple machines, which is key for running huge AI models.
  

• Open for Everyone: Meta shared these improvements as open-source code, so developers can build on them easily.
  

• Clear Speed Gains: Tests show big speed boosts on cutting-edge GPUs, especially for long and complex AI tasks.
  

Older methods like padding or slicing forced the AI to process extra empty or awkward data chunks, which slowed everything down. MetaShuffling’s approach avoids that by reorganizing data dynamically, letting the AI use its resources more fully and run faster. It’s a smarter, leaner way to handle big AI models -  making them more practical for real-world use. (source)

🚀 Google Launches AI Futures Fund for Startups

Google just launched the AI Futures Fund to boost early-stage startups building the future of AI. The program offers early access to advanced models like Gemini, Veo, and Imagen, direct collaboration with Google DeepMind and Google Labs experts, equity investment opportunities, plus Google Cloud credits and dedicated support to help startups grow and scale their AI products.

Apply here!

The Tools

I checked out Zero this week -  an AI-native email client that manages your inbox so you don’t have to lift a finger.

What It Does
Zero uses AI to prioritize your emails, summarize long threads, draft replies, and organize your inbox into smart categories. It even lets you chat directly with your email, all while understanding the full context of your workflow. Unlike traditional email clients, Zero acts like a proactive assistant that keeps you ahead of your inbox chaos.

How to Use It:
Step 1: Sign up for the public beta at 0.email.

Step 2: Connect your email account and let Zero start prioritizing, summarizing, and drafting replies automatically.

Step 3: Chat with your inbox naturally to plan your day, manage projects, or get quick updates- right from your email client.

The Algorithm

• Hugging Face drops an epic deep-dive on Vision-Language Models (VLMs): From agentic VLMs and MoEs to video understanding and alignment- this blog maps out where the field is and where it's heading. Curious about VLMs or building with them? Start here.

• Similarweb’s new Global AI Tracker shows GenAI traffic is booming: ChatGPT is back on the rise, Grok and Gemini are gaining steam, and DeepSeek’s momentum is holding strong. If you’re watching the GenAI race, this dashboard is a pure signal.

• Meta’s CATransformers is a pioneering AI framework that cuts carbon emissions by ~9.1% while maintaining performance, using a first-ever co-design of neural architectures and hardware. Powered by the Architectural Carbon Tool (ACT), it pushes AI sustainability forward- check out the paper and code in Meta’s Sustainable AI hub.

• Last week, we launched more free AI & data courses to help you sharpen your technical skills and explore powerful tools in practice- 

  • Build Data Pipelines with Apache Airflow: Learn workflow orchestration by building real-world ETL pipelines using Apache Airflow. From mastering the basics of DAGs and task dependencies to advanced scheduling, this course guides you through using PythonOperators, BashOperators, cron expressions, and hooks. Perfect for anyone wanting scalable, clean, and reproducible data engineering pipelines.

  • Getting Started with Tableau: Jump into data visualization with Tableau, the industry-leading tool. This beginner-friendly course covers connecting data sources, creating compelling visualizations, and building interactive dashboards. Gain essential skills like data prep, calculated fields, and sharing insights- all through Tableau’s intuitive interface.

The Output

From cracking complex math to making healthcare chatbots actually trustworthy, this week showed AI getting seriously practical. Whether it’s Google’s AlphaEvolve cooking up new algorithms, OpenAI setting higher bars for medical AI, or CodeRabbit making coding less stressful-  the focus is shifting from flashy demos to real-deal tools that help you get stuff done.

Feels like AI is settling into the everyday grind, and honestly, that’s pretty exciting. 

See you next week! 👋