Good morning. It’s Wednesday, May 21st.

On this day in tech history: 1952: IBM announced its first electronic computer, the Model 701

In today’s email:

• Google I/O 2025

• Microsoft Build 2025

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Google I/O 2025 Recap: Multimodal AI Takes Center Stage with Smarter Search and Generative Tools

Google’s latest AI advancements introduced at I/O 2025 include a range of multimodal models and tools grouped by functionality.

The Gemini family leads with Gemini 2.5 Pro “Deep Think,” which offers parallel hypothesis evaluation and configurable thinking budgets for complex math and coding, surpassing competitor benchmarks. Gemini 2.5 Flash enhances efficiency with thought summaries and reduced token use, available now on AI Studio, Vertex AI, and Gemini app. Gemini 3n targets smartphones and edge devices with lightweight multimodal processing and advanced multilingual support. Gemini Diffusion accelerates text generation through iterative noise refinement, while MedGemma and SignGemma focus on medical image analysis and sign language translation, respectively.

In AI storytelling and filmmaking, Veo 3 delivers lifelike video generation with integrated audio, lip syncing, and physics, accessible via Gemini, Flow, and Vertex AI for Ultra subscribers. Imagen 4 produces high-quality 2K images with improved texture detail and typography, supporting multiple platforms including Gemini and Whisk. Flow combines Veo 3, Imagen, and Gemini to enable cinematic scene creation with story management. Lyria RealTime offers interactive music generation via Gemini API, supporting detailed control over genre, instruments, and audio quality.

For coding and developer tools, Agentic Colab autonomously analyses and refines code across notebook cells with error fixes and data workflows. Gemini Code Assist 2.5 provides AI co-pilot and multi-file code review capabilities, free for individuals and GitHub users. Firebase Studio converts Figma designs into full-stack apps with backend provisioning, supported by Firebase AI Logic integrating Gemini and Vertex AI. Jules, powered by Gemini 2.5 Pro, is an asynchronous coding agent running in a secure Google Cloud VM for bug fixes, testing, and prototyping, including audio changelogs. Stitch transforms text and images into UI designs with frontend code, supporting Figma export and theming. Google AI Studio now integrates Gemini 2.5 Pro, Imagen 4, and Veo 3 within its code editor, linked to the GenAI SDK for instant app generation.

The Gemini API introduces native audio output, live and async calls, and smart video/audio features with text-to-speech voice customization. New AI Mode on Gemini 2.5 improves question breakdown for research, live camera queries, and data charting, now available in the U.S. Gemini in Chrome integrates AI directly in the browser for summarization and task automation, initially to U.S. users.

Additional innovations include Google Beam, a 3D video communication platform converting 2D streams into volumetric experiences with light field displays; Project Astra, an experimental AI assistant offering proactive task management via contextual device data; and Google Meet’s real-time speech translation using a DeepMind audio model preserving speaker tone, initially supporting English and Spanish. Google AI Ultra, a new subscription tier available in the U.S. for $249.99/month, bundles access to these advanced models and tools, including Gemini, Flow, Whisk, NotebookLM with expanded usage, plus 30 TB storage and YouTube Premium.

Sergey Brin also made an unannounced appearance, explaining that the promise of AGI pulled him out of retirement. He remarked, “Anyone who is a computer scientist should not be retired right now,” underscoring the rapid pace of AI advances. Brin also speculated that our world might exist within “a stack of simulations.”

Construction Underway for First Stargate AI Data Center in Texas

Microsoft Build 2025: Agents That Code, Coordinate, and Accelerate Discovery

At Build 2025, Microsoft laid out its vision for an “open agentic web,” unveiling over 50 AI updates across software development, system architecture, and scientific research. GitHub Copilot now handles asynchronous tasks like drafting pull requests, refactoring code, and fixing bugs under human review. Agent coordination is handled via Azure AI Foundry using Agent2Agent (A2A) and Anthropic’s Model Context Protocol (MCP), which enables shared memory and multi-agent fluency. MCP is now embedded directly into Windows, giving agents standardized access to directories, subsystems, and built-in permissions to safeguard misuse, demonstrated with Perplexity navigating a local file system via natural language.

Edge compute gets a boost through Windows AI Foundry, allowing on-device agent execution across Intel, AMD, Nvidia, and Qualcomm chips, no custom runtime required. Developers can fine-tune foundation models with proprietary data via Copilot Tuning, deploy Grok 3, or choose from over 1,900 models on Azure. NLWeb, Microsoft’s new protocol, introduces personalized AI agents to websites without relying on the cloud.

A standout demo featured Microsoft Discovery, an AI-driven research engine that acts like a virtual team of scientific “postdocs.” It screened 367,000 chemical candidates in just 200 hours to identify a safer coolant for data centre immersion—a process that traditionally takes years.

Built on a graph-based knowledge engine with Copilot integration, Discovery allows researchers to simulate, test, and explore scientific hypotheses through natural language. Microsoft plans to extend its use across pharma, semiconductors, and cosmetics, with integration into Nvidia hardware, Synopsys tools, and eventually quantum computing.

• xAI and Tesla Build First Gigawatt AI Training Supercluster

• Intel is reportedly exploring a sale for its networking and edge unit

• The Netherlands is building a leading neuromorphic computing industry

• NVIDIA Unveils Llama Nemotron Nano 4B Model for AI Edge

• Stable Video 4D 2.0: New Upgrades for High-Fidelity Novel-Views and 4D Generation from a Single Video

• Captive to industry, robots now dream of work, with no electric sheep in sight

• Duke's robot dog mimics human touch, sound to navigate forest terrain

• AI in the workplace is nearly 3 times more likely to take a woman’s job as a man’s, UN report finds

• Report: Apple plans to make its large language models available to developers

• OpenAI's planned data center in Abu Dhabi would be bigger than Monaco

• Large language models often struggle with decision-making — a new study explains why

• US team develops world's fastest quantum switch to supercharge AI by 1 million times

• Watch: Co-Founder of Neuralink Max Hodak says AGI & ASI by 2030-2035

• Trump signs the Take It Down Act into law

• A comprehensive guide to top humanoid robot builders

• AI goes to 'kindergarten' in order to learn more complex tasks

• Workday adds seven new AI agents to its cloud platform

• AI Can Beat You in a Debate When It Knows Who You Are, Study Finds

• World's first computer that combines human brain with silicon now available

📄 Chain-of-Model Learning for Language Model

📄 AdaptThink: Reasoning Models Can Learn When to Think ​​

📄 AdaCoT: Pareto-Optimal Adaptive Chain-of-Thought Triggering via Reinforcement Learning

📄 Emerging Properties in Unified Multimodal Pretraining

📄 SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!