Good morning. It’s Friday, May 30th.

On this day in tech history: 1992: Apple announced the Newton, a personal digital assistant (PDA) that was ahead of its time, featuring handwriting recognition and early mobile computing capabilities. Though it was not commercially successful, it laid the groundwork for future devices like the iPhone.

In today’s email:

• Meta + Anduril Partnership

• DeepSeek Upgrade

• HuggingFace Open-Source Robots

• Odyssey 3D World Generator

• 5 New AI Tools

• Latest AI Research Papers

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

Meta’s AI Hits 1 Billion Users, Moves into Military Tech

Meta’s AI ambitions now extend across scale and strategy, with its assistant quietly crossing 1 billion monthly users across WhatsApp, Instagram, Facebook, and Messenger. This milestone follows the launch of a standalone AI app and a roadmap that leans into voice, personalized help, and entertainment. While the service remains free, Mark Zuckerberg has flagged possible monetization through compute-heavy subscriptions or premium recommendations, noting user behavior will guide the approach.

On the research front, Meta’s FAIR team and The Hebrew University overturned a core assumption about large language models: longer reasoning chains don’t mean better answers. Their study showed shorter chains improve accuracy by up to 34.5%, slash compute by 40%, and cut inference time by a third. The “short-m@k” method runs parallel short reasoning paths and uses majority voting to pick the best result, outperforming traditional chain-of-thought techniques. Training on short sequences boosts performance, while fine-tuning on longer ones adds delay without payoff.

Beyond the consumer and research layers, Meta is extending its AI capabilities into defense through a partnership with Anduril, the military-tech firm founded by Oculus creator Palmer Luckey. The two companies are combining Meta’s AR infrastructure with Anduril’s autonomous battlefield systems, designed to enhance situational awareness without increasing cognitive overload. That means less fog of war, more real-time clarity. Palmer Luckey, founder of Anduril and Oculus, claims the collaboration could reduce both casualties and defense costs.

Unlike legacy contractors, Anduril develops and funds its own platforms, touting speed and flexibility over red tape. It’s an approach that sets it apart from incumbents like Lockheed and Boeing. Read more.

China’s DeepSeek Gets Major Upgrade

DeepSeek’s latest upgrade, R1-0528, puts its open-source reasoning AI firmly in the ring with OpenAI’s o3 and Google Gemini 2.5 Pro. This update ramps accuracy on tough tests like AIME 2025 from 70% to 87.5%, and coding chops on LiveCodeBench jumped nearly 10%. Behind the scenes, smarter algorithms and beefed-up compute let it chew through longer token sequences, up to 23,000 tokens per question. The model now speaks JSON, handles function calls, and drops the awkward “thinking mode” token, making developers’ lives a little easier.

For those running on leaner hardware, a distilled 8-billion-parameter variant plays nice with GPUs around 16GB VRAM. The open MIT License stays intact, and API pricing is refreshingly transparent, starting at $0.14 per million input tokens during busy hours. Early buzz praises its coding smarts and edges it closer to the big-name competitors, proving open source can pack a punch. Read more.

Hugging Face Unveils Open Source Humanoid Robots

Hugging Face is stepping beyond AI models and into the hardware aisle with two new open-source humanoid robots—HopeJR and Reachy Mini. HopeJR is a full-body unit with 66 degrees of freedom and the ability to walk, gesture, and grasp, all for under $3,000. Reachy Mini, on the other hand, is a desktop companion designed to test voice-based and interactive AI apps, priced around $250–$300.

Built in collaboration with The Robot Studio and powered by knowledge gained from Hugging Face’s acquisition of Pollen Robotics, both bots are designed to be hackable, rebuildable, and affordable. Developers can modify hardware and software alike, with builds supported by Hugging Face’s LeRobot ecosystem of open datasets and model libraries. No ship date yet, but early access is open, and the company expects initial units to roll out by year’s end. Read more.

Odyssey, backed by Pixar co-founder, now streams AI-Generated 3D worlds with real-time navigation

Odyssey, backed by Pixar’s co-founder and led by former Cruise and Voyage executives Oliver Cameron and Jeff Hawke has introduced a prototype AI world model that streams AI-generated 3D environments in real time. The system renders 360-degree video frames every 40 milliseconds, enabling users to navigate immersive spaces with simple directional commands, no traditional game engine required.

Drawing on data from a bespoke backpack-mounted camera rig, Odyssey’s model emphasizes spatial consistency and temporal coherence, maintaining stable visuals for over five minutes. Running at 30 frames per second on Nvidia H100 clusters, the current cost hovers between $1–$2 per user-hour. Unlike DeepMind or Microsoft, Odyssey focuses on open-ended, interactive worlds with richer video-driven dynamics. Read more.

• Mistral's New Embedding Model Tops OpenAI and Cohere in Code Retrieval Tests

• Factory’s ‘Droids’ AI agents take on the full software development cycle

  Perplexity’s Labs gets serious about long-haul AI tasks

• Quantum computers could break Bitcoin encryption 20 times faster, Google warns

• GPT-4o Mimics Human Cognitive Dissonance, Study Finds

• Scientists create robots that take their first steps straight out of the 3D printer

• Amazon and The New York Times Announce an A.I. Licensing Deal

• ManusAI Launches Manus ‘Slides’

• Black Forest Labs' FLUX.1 merges text-to-image generation with image editing in one model

• Nvidia, Dell to supply next US Department of Energy supercomputer

• AI outperforms 90% of human teams in a recent hacking competition with 18,000 participants

• Don’t Mistake AI’s Text Trails for Thinking, Says New Research

• xAI to pay Telegram $300M to integrate Grok into the chat app

• Anthropic CEO Flags AI’s Potential to Halve Entry-Level White-Collar Jobs by 2030

• Netflix co-founder Reed Hastings joins Anthropic’s board of directors

• 68% of tech vendor customer support to be handled by AI by 2028, says Cisco report

• Writer becomes one of the first AI model developers to complete ISO 'trust triad'

• Atlas gets real-world upgrade with superhuman vision and near-perfect object tracking

• Rick Rubin Calls Vibe Coding the Punk Rock of AI, Awaits Authenticity

• Anti-Aging Cocktail Extends Mouse Lifespan by About 30 Percent

• A new transformer architecture emulates imagination and higher-level human mental states

📄 The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models

📄 ScienceBoard: Evaluating Multimodal Autonomous Agents in Realistic Scientific Workflows ​​

📄 R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing

📄 Sherlock: Self-Correcting Reasoning in Vision-Language Models

📄 ZeroGUI: Automating Online GUI Learning at Zero Human Cost  

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!