In partnership with

Good morning. It’s Monday, May 19th.

On this day in tech history: 2001: Apple inaugurated its first two retail locations in McLean, Virginia, and Washington, D.C. These stores attracted 7,700 visitors and generated nearly $600,000 in sales during their opening weekend.

In today’s email:

• Nvidia’s Full-Stack Compute Push

• GPT-5 to Enhance Tool Cohesion

• Grok Now Generates Custom Image Ratios and On-Demand Charts

• 3 New AI Tools

• Latest AI Research Papers

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

GPT-5 to Enhance Tool Cohesion, Reduce Model Switching

OpenAI confirms GPT-5 will prioritize improving integration and reducing model switching between components like Codex, Operator, and Memory rather than introducing a new architecture. Initially aiming to unify GPT and “o” models, technical hurdles led to separate releases of o3 and o4-mini reasoning models. Operator, the screen agent, will soon receive updates for increased reliability. Token usage is expected to rise continuously, reflecting expanded real-world application rather than fixed benchmark improvements.

Watch: Altman Proposes ChatGPT as a Continuous Life Recorder

OpenAI CEO Sam Altman envisions ChatGPT evolving into a persistent digital assistant that continuously stores a user’s entire life context—conversations, emails, reading materials, and more—amounting to “trillion tokens of context.” Younger users treat ChatGPT as a life advisor for complex decision-making, while older users view it as a Google substitute. Privacy and trust remain critical concerns given the scale of personal data involved and the risks of bias or misuse.

Video starts at: 26:53

Inside Nvidia’s Full-Stack Compute Push

Walking through Nvidia’s recent announcements, you get the sense they aren’t just building faster chips. They’re reshaping the entire landscape of computing.

Start with quantum. Nvidia and Japan’s AIST just launched ABCI-Q, a supercomputer stitched together from more than two thousand H100 GPUs. It runs on Nvidia’s CUDA-Q platform and is connected by Quantum-2 InfiniBand. But the real story is what it blends together: superconducting chips from Fujitsu, neutral atom tech from QuEra, and photonics from OptQC. It’s a hybrid system built to scale quantum-classical simulations, the kind of infrastructure that could unlock major breakthroughs in healthcare and energy. It feels like a preview of the future of scientific research, only it’s already running.

Then came Computex. Among all the announcements, one stood out. Nvidia introduced Isaac GR00T N1.5, a robot model trained using synthetic motion created from still images. They call the method GR00T-Dream. It skips human demonstrations entirely and cuts training time from months to just 36 hours. The model is built for Jetson Thor and already shows better task performance in industrial robotics. Watching it in action, you get a sense of how fast this kind of system could spread into real-world applications.

Of course, none of this works without serious infrastructure. Nvidia revealed NVLink Fusion, which creates an 800 gigabit per second connection between CPUs and GPUs. It supports silicon from companies like MediaTek and Astera Labs, allowing for modular AI-native compute clusters. The result is a flexible and insanely fast platform that makes no distinction between types of chips, only whether the job gets done.

For those who want this kind of power on-site, Nvidia’s new DGX Spark and DGX Station systems bring supercomputer performance into the office. The Spark runs on the GB10 Superchip and delivers one petaflop of compute. The Station steps it up with the GB300 Ultra, offering 20 petaflops and nearly 800 gigabytes of memory. Both systems are designed for model fine-tuning, slicing with multi-instance GPUs, and running inference without touching the cloud. They are essentially portable AI data centers.

Nvidia is also pushing further into simulation with its Omniverse Blueprint tool. It lets engineers design and test entire AI factories in 3D before a single part is built. These simulations use OpenUSD and SimReady assets, and integrate with platforms from Siemens, Delta, and Jacobs. Everything from cooling and power to workflow and networking can be modeled and optimized against Nvidia’s GB200 NVL72 systems. It’s not just virtual design. It’s operational rehearsal for the future.

And behind all this, Nvidia is making a long bet on quantum. The company is reportedly in talks to invest in PsiQuantum’s $750 million raise. It’s a move that fits Nvidia’s broader strategy of combining classical and quantum compute. Jensen Huang captured the mindset clearly when he said, “I’ll hire one biological engineer, then rent a thousand AIs.” That line isn’t just clever. It’s the entire playbook.

Grok Now Generates Custom Image Ratios and On-Demand Charts

Musk’s xAI has upgraded its Grok AI assistant with two new capabilities: visual chart generation and customizable image aspect ratios. Chart creation is currently available on the browser, with plans to expand across platforms soon. Users can now generate images tailored to their preferred dimensions directly within Grok. Read more.

• Chinese Robot Combat Wars To Air on TV

• Soft Robotics, No CPU Required

• AWS Open-Sources Strands Agents SDK to Speed AI Agent Deployment

• Stability AI Open-Sources Compact Mobile Text-to-Audio Model

• Y Combinator startup Firecrawl is ready to pay $1M to hire three AI agents as employees

• Mark Gurman: Apple is “unlikely” to spend much time discussing Siri, and is preparing to separate the Apple Intelligence brand from Siri in marketing

• Microsoft wants its AI Copilot app to lure Gen Z from rivals by behaving like a therapist

• China rolls out fleet of driverless mining trucks

• OpenAI Releases Addendum on Codex

• MIT disavows doctoral student paper on AI's productivity benefits

• With AI, researchers predict the location of virtually any protein within a human cell

• Anthropic is forced to apologize after Claude undercuts its legal team

• AI study finds 'no significant impact on earnings or recorded hours'

• China begins assembling its supercomputer in space

• AI chips are the new 'coin of the realm' as they grease the wheels of geopolitical negotiations, BofA says

• World's first AI doctor clinic now open in Saudi Arabia

• AI Chatbots Mirror a Human Brain Disorder

📄 Qwen3 Technical Report

📄 Visual Planning: Let's Think Only with Images ​​

📄 Parallel Scaling Law for Language Models

📄 OpenThinkIMG: Learning to Think with Images via Visual Tool Reinforcement Learning

📄 Group Think: Multiple Concurrent Reasoning Agents Collaborating at Token Level Granularity

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!