In partnership with

Good morning. It’s Wednesday, March 14th.

Did you know: 1643: The earliest known adding machine, a mechanical calculator called Pascaline, was invented by 19-year-old Blaise Pascal. It could add, subtract, and handle carries, marking a significant step in computing history.

In today’s email:

• Google I/O

• US-Saudi Forum Tech Investment

• OpenAI’s New Features

• 5 New AI Tools

• Latest AI Research Papers

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

Ahead of Google I/O, Tech Giant Teases “Gemini Everywhere”

Google’s Keynote Event, I/O 2025 will run May 20–21, with Sundar Pichai and Demis Hassabis leading presentations. Expect Gemini Everywhere: powering search, running in Chrome, embedded in Android XR, and demoed in Waymo’s self-driving cars. A separate Android event airs a week earlier, setting the stage for Google’s next AI chapter.

Gemini Is Coming for Developers: Ahead of I/O, Google teased a new AI agent for developers. It's built to handle documentation, bug detection, and task tracking and it’s getting voice integration through XR headsets. Think: hands-free coding sessions powered by Gemini, with deeper demos expected on stage next week.

From Watches to Windshields: Gemini is replacing Google Assistant across Wear OS, Android Auto, Google TV, and XR. On smartwatches, it’s your reminder-setter and quick-reply tool. In cars, it handles routing, summaries, and follow-up questions mid-drive. On Google TV, it helps you pick a movie and explains Roman history. Processing stays in the cloud for now, but Google’s working with automakers to run it on local hardware.

AI Mode Tests in Search: A rainbow-bordered “AI Mode” is popping up in place of the classic “I’m Feeling Lucky” button. It’s part of Google’s larger bet on generative answers and predictive search. Also coming: an agent that flags bugs and security issues during software builds—another nudge toward AI-assisted engineering.

Visual Shopping Meets Veo AI: Shopping on Google just got a 3D upgrade. With three images, a new system powered by Veo creates a rotating product view—for everything from sofas to sneakers. Google’s also expected to roll out a Pinterest-style tool for saving fashion and home inspo, wrapped in a commercial search strategy that keeps ad dollars close.

Backing the Next Gen of AI Startups: Google’s AI Futures Fund is now live, offering DeepMind access, Cloud credits, and cash to early-stage startups—no deadline, just traction. Participants already include Viggle and Toonsutra. It’s part of a growing portfolio that includes a $20M Google.org effort and a $120M Global AI Opportunity fund.

U.S.–Saudi AI Forum Sets Stage for Global Tech Realignment

Saudi Arabia has launched a new AI company called Humain, backed by the $700B Public Investment Fund and chaired by Crown Prince Mohammed bin Salman. The focus: Arabic-language models, massive local data centers, and plugging the Kingdom deep into the global AI supply chain. It complements prior initiatives like Alat, which targets $100 billion in AI hardware investment by 2030 as Riyadh bids to become a global AI hub.

The U.S. is scrapping blanket chip export controls for Gulf states. Instead, it’s signing one-on-one tech security deals—starting with Saudi Arabia and the UAE. That opened the door for Nvidia and AMD to sell top-tier AI chips to firms like Humain and UAE-based G42. Sam Altman helped grease the wheels, and now U.S. firms like Microsoft and OpenAI are doubling down on the region.

At the Saudi-U.S. Investment Forum, Nvidia CEO Jensen Huang confirmed Saudi Arabia will get 18,000 of its latest Blackwell GPUs as part of a $10B plan to build 500 megawatts of AI capacity. AMD’s in the mix too. It’s just the first phase as Humain wants hundreds of thousands of chips to power future genAI and Arabic-language models. Nvidia shares jumped 5% on the news.

Amazon Web Services is going deep, investing $5B+ in a new “AI Zone” in partnership with Humain. The deal includes cloud infra, training hubs, and startup support all built within Saudi borders to comply with strict data localisation laws. It’s separate from AWS’s planned $5.3B regional expansion, meaning Saudi is now a core node in Amazon’s global AI map.

OpenAI tightens enterprise grip with new tools and AGI signals

Clinical AI Gets a Gold Standard: OpenAI has launched HealthBench, a benchmark built with input from 262 physicians to evaluate LLMs in clinical use across 5,000 multi-turn conversations. Models are scored on 48,000 expert-written criteria spanning accuracy, clarity, and instruction-following. GPT-4.1 nano outperformed GPT-4o despite being 25x cheaper, highlighting trade-offs in speed and cost. The benchmark is open-source via simple-evals on GitHub.

Enterprise Tools Get an Upgrade: A new PDF export feature in Deep Research now allows Plus, Team, and Pro users to download formatted reports with tables, images, and clickable citations intact. It answers growing demand for verifiable, distributable outputs and positions. ChatGPT now integrates with Microsoft SharePoint, enabling retrieval and summarization of internal documents. Users can blend proprietary data with public sources to generate reports, all while preserving access controls.

AI Research, Then Revenue: Chief Scientist Jakub Pachocki notes that OpenAI’s models are beginning to generate novel insights, driven by unsupervised pre-training and RLHF. He sees AI moving from reasoning toward economic impact like autonomous software development and valuable research outputs by year-end. CEO Sam Altman, in an interview with Sequoia Capital forecasts 2025 as the year of working AI agents, 2026 as a tipping point for scientific breakthroughs, and 2027 as the dawn of economically useful robotics.

• xAI’s Colossus Supercomputer Scales Up

• Tesla’s Optimus robot trains fully in simulation, shows real-world progress

• Watch: Elon Musk: Universal “High” Income

• Watch: Google's Chief Scientist Jeff Dean says we're a year away from AIs working 24/7 at the level of junior engineers

• New Anthropic model “claude-neptune” enters safety evaluation

• Microsoft releases ARTIST, a reinforcement learning framework for LLMs with agentic tool use

• Lloyd’s insurers launch first AI chatbot error policies

• Kling 2.0 edges ahead of Veo 2, Runway Gen-4 in image-to-video benchmarks

• TikTok launches TikTok AI Alive, a new image-to-video tool

• Spotify AI DJ introduces voice command personalization

• Notion takes on AI notetakers like Granola with its own transcription feature

• MIT physicists snap the first images of “free-range” atoms

• xAI's promised safety report is MIA

• Tencent hires WizardLM team, a Microsoft AI group with an odd history

• Microsoft slashes 6,000 jobs worldwide in strategic shift toward AI

• Apple to unleash AI-powered battery saving update for iPhones in iOS 19

• Epoch AI Predicts Slowdown for Reasoning Models by 2026

• Manus AI has officially launched publicly

• LegoGPT generates buildable Lego models from text descriptions

• Freepik launches F Lite 7B, a faster, distilled model under open license

• Perplexity eyes $14 billion in valuation in fresh funding round, WSJ reports

• Nations meet at UN for 'killer robot' talks as regulation lags

• Pika's Ad Positions AI as Post-Apocalyptic Escape

• Trump fires head of Copyright Office two days following report that AI training may not be fair use

📄 Seed1.5-VL Technical Report 

📄MiMo: Unlocking the Reasoning Potential of Language Model -- From Pretraining to Post Training ​​

📄 Step1X-3D: Towards High-Fidelity and Controllable Generation of Textured 3D Assets

📄 Learning from Peers in Reasoning Models

📄 Unified Continuous Generative Models

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!