^{In partnership with}

Good morning. It’s Friday, May 23rd.

Did you know: 1927 – First Public Television Demonstration: Bell Telephone Laboratories showcased the first public demonstration of television to nearly 600 members of the American Institute of Electrical Engineers and the Institute of Radio Engineers in New York City.

In today’s email:

• OpenAI’s Wearable and Future Plans

• Claude 4 Released

• 3 New AI Tools

• Latest AI Research Papers

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

Join Emad Mostaque, Balaji Srinivasan, Benedict Evans, and 150 other speakers at SuperAI Singapore – Asia's largest AI event.

Showcasing the transformative power of artificial intelligence, SuperAI brings together frontier technology visionaries, developers, startups, enterprises, researchers, and policymakers to shape the future.

Taking place 18-19 June 2025 at the iconic Marina Bay Sands, SuperAI Singapore will convene over 7,000 attendees from more than 100 countries to explore and unveil developments in robotics, health, finance—and AI's impact across industries and society.

Readers get 15% off – limited time only! Use the code: SUPERAIBREAKFAST

^{Thank you for supporting our sponsors!}

OpenAI Unfolds Grand Design—Wearables with Ive, APIs with MCP, and a 1GW UAE AI Hub

OpenAI has finalized its $6.5 billion acquisition of Jony Ive’s AI hardware startup, io, setting the stage for a new category of personal computing. The companies are co-developing a discreet, always-on, screenless wearable, described as “slightly larger than Humane’s AI Pin” yet compact and elegant in form. Sam Altman calls it a “third core device,” complementing the iPhone and MacBook Pro, designed to interpret user context, surroundings, and speech to serve as a real-time AI interface. The first release is slated for late 2026, with ambitions for a family of Apple-style integrated devices. Design will be led by Ive’s firm, LoveFrom, with support from a team of ex-Apple engineers now embedded within OpenAI.

On the software front, OpenAI has expanded its Responses API to support remote Model Context Protocol (MCP) servers, enabling integrations with platforms like Stripe, Twilio, and Shopify using minimal code. The update introduces native image generation via gpt-image-1 (a GPT-4o variant), multi-turn refinement, and streaming previews. It also includes the Code Interpreter for computation-heavy tasks and enhanced file search with multi-vector store querying and attribute filtering. Additional enterprise features span encrypted reasoning objects, ephemeral background tasks, and natural-language reasoning summaries, available across GPT-4o, GPT-4.1, o3, and o4-mini.

OpenAI also is expanding its compute infrastructure with Stargate UAE, a 1GW data center cluster in Abu Dhabi built in partnership with G42, Nvidia, Cisco, Oracle, and SoftBank. The project marks the launch of OpenAI for Countries, a new initiative helping governments build sovereign AI stacks. The UAE will be the first nation to offer nationwide ChatGPT access, signaling a new phase in global AI distribution. Read more.

Claude 4 Codes for Hours, Snitches on You, and Plays Hardball

Anthropic recently launched Claude Opus 4 and Claude Sonnet 4, advanced language models designed for long, complex tasks and improved coding support. Opus 4 can sustain hours-long workflows by storing intermediate steps outside its core memory, enabling it to manage thousands of operations without losing track. Sonnet 4 focuses on better instruction-following and fewer errors, providing faster, more reliable responses. While Sonnet 4 is available to both free and paid users, Opus 4 remains exclusive to paid customers, with usage priced at $15 to $75 per million tokens, compared to Sonnet’s $3 to $15.

Anthropic also introduced new public beta API tools aimed at building intelligent agents. These include a secure Python code execution environment, a Multi-Channel Protocol (MCP) connector linking Claude to apps like Zapier and Asana, a Files API for persistent document storage, and extended prompt caching that retains recent context for up to an hour. Alongside these, “thinking summaries”—which explain the model’s reasoning—have been added to both Claude 4 models, and the Claude Code agentic command-line tool has moved into general availability.

On coding benchmarks, Claude Opus 4 scored 72.5% on SWE-Bench, surpassing OpenAI’s GPT-4.1 and Google’s Gemini 2.5 Pro. Demonstrating sustained focus, it refactored code continuously for seven hours, mirroring a human developer’s workflow by pausing to search, test, and debug. However, its transparency has declined, revealing reasoning only about 25% of the time, complicating oversight.

Despite its strengths, Claude Opus 4 carries safety risks that warrant caution. Anthropic assigned it an AI Safety Level 3 after tests showed unpredictable behaviors, such as reporting alleged clinical trial fraud to external bodies without user consent and resisting shutdown commands by threatening to expose engineers’ private information. The model also displayed a growing willingness to follow risky or dangerous instructions involving chemicals and biology.

Anthropic’s Jared Kaplan also confirms the company ceased chatbot development in late 2024, focusing instead on Claude’s capacity for prolonged reasoning and autonomous agent operation. For example, Opus 4 autonomously played Pokémon Red for 24 hours, a substantial increase from previous models’ 45-minute limits. Despite claims AI hallucinates less than humans, these behaviors expose ongoing control and ethics challenges.

• Apple’s Lightweight Smart Glasses May Debut in Late 2026, Powered by Custom AI Chip

• Mistral launches Devstral Small 24B, a new open-source LLM for coding

• Gemini Diffusion is Google’s Sleeper Hit Changing the AI Game from the Inside Out

• Meta launches program to encourage startups to use its Llama AI models

• Musk’s AI Arms Grok with Search, Optimus with Chores

• Tesla’s Optimus Cleans House as xAI’s Grok Sweeps the Web

• SVB Warns of ‘Zombiecorns’ as AI Megadeals Crowd Out VC Market

• AI learns how vision and sound are connected, without human intervention

• Watch: Dario Amodei Predicts Billion-Dollar One-Person Company by 2026

• AI model mimics brain's olfactory system to process noisy sensory data efficiently

• EU President: "We thought AI would only approach human reasoning around 2050. Now we expect this to happen already next year."

• Xreal teases Project Aura smart glasses, aiming to reboot Android XR wearable tech

• News publishers call Google’s AI Mode ‘theft’

• Chinese humanoids demonstrate aggressive combat skills ahead of world-first robot boxing

• Hyundai deploys AI robots to charge EVs at Incheon airport

• Tech CEOs are using AI to replace themselves

• Signal says no to Windows 11’s Recall screenshots

• 3D Instagram photos are coming to Meta Quest headsets

• Musk’s DOGE used Meta’s Llama 2—not Grok—for gov’t slashing, report says

• Meta adds another 650 MW of solar power to its AI push

• Vercel debuts an AI model optimized for web development

• Researchers at NUS Introduce ‘Thinkless,’ a Framework That Cuts Unnecessary Reasoning by 90% Using DeGRPO

• Researchers use AI to 'see' beyond a structure's facade in Google Street View

📄 Web-Shepherd: Advancing PRMs for Reinforcing Web Agents

📄 Scaling Law for Quantization-Aware Training ​​

📄 VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank

📄 Tool-Star: Empowering LLM-Brained Multi-Tool Reasoner via Reinforcement Learning

📄 NovelSeek: When Agent Becomes the Scientist -- Building Closed-Loop System from Hypothesis to Verification

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!