In partnership with

Good morning. It’s Wednesday, October 29th.

On this day in tech history: In 2014, Google quietly released word2vec, the neural embedding model that redefined how machines understand language. Built by Tomas Mikolov’s team, it compressed linguistic meaning into vectors, birthing modern NLP and powering everything from GPT tokenizers to semantic search. Every “smart” autocomplete today traces back to that moment in a Google Brain repo.

In today’s email:

• OpenAI Becomes a For-Profit

• NVIDIA’s Latest Power Play

• Microsoft turns office workers into app builders

• 5 New AI Tools

• Latest AI Research Papers

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

OpenAI cements for-profit shift and expert-verified AGI framework in new Microsoft deal

OpenAI is now officially a for-profit public benefit corporation, with its nonprofit parent retaining a 26–27 percent stake to oversee research priorities. The restructure, approved by California and Delaware regulators, frees OpenAI to raise capital without limits and chase big, audacious AI projects, including the pursuit of artificial general intelligence (AGI).

Microsoft holds a 27 percent stake, keeps IP rights to OpenAI’s models through 2032, and can independently pursue AGI, but OpenAI can no longer simply declare AGI on its own anymore. Any claim must be verified by an independent expert panel first. Both companies can still chase AGI independently, and Microsoft can team up with other partners, though compute thresholds are in place if it uses OpenAI’s IP.

Watch the livestream replay here.

Infrastructure is at a staggering scale. OpenAI has already committed $1.4 trillion to roughly 30 gigawatts of data center capacity and aims to scale up to a gigawatt per week at $20 billion per gigawatt. This scale underpins OpenAI’s next big goal: a fully autonomous “legitimate AI researcher” by 2028, following an intern-level research assistant milestone in 2026. Jakub Pachocki, chief scientist, explains these systems will leverage expanded “test time compute,” letting models dedicate entire data centers’ worth of processing power to solve complex problems.

Meanwhile, OpenAI now flags that over 500,000 ChatGPT users weekly may show signs of mania or psychosis, with millions more relying heavily on the chatbot for emotional support. To address this, GPT-5 was retrained with guidance from 170 psychiatrists and clinicians. The latest update, “gpt-5-oct-3,” adds clinical taxonomies, hotline redirects, and session-break nudges to detect distress without acting as a therapist.

Power is now part of the equation. In a formal pitch to the White House, OpenAI warned that America’s AI ambitions hinge on energy, bluntly calling electricity “the new oil.” Its Stargate data centers alone will demand 10 GW, enough to power eight million homes, and the company is pushing for 100 GW per year nationally to support next-gen AI workloads. Read more.

Inside Nvidia’s biggest power grab yet: AI factories, quantum leaps, and America’s tech backbone

Nvidia just made its biggest pitch yet to own the future of computing. At GTC DC 2025, Jensen Huang laid out an aggressive plan, all designed to cement Nvidia as the backbone of America’s tech infrastructure.

The centerpiece is the “AI factory,” massive data centers built with the Department of Energy and Oracle to simulate and train industrial-scale AI systems. The first, Equinox, will pack 10,000 Blackwell GPUs; its follow-up, Solstice, targets 2,200 exaflops of compute - orders of magnitude beyond today’s AI clusters. These facilities run on Nvidia’s Omniverse DSX, which lets engineers model gigawatt-scale campuses as physics-accurate digital twins before they’re even built.

Nvidia is also bringing its fastest AI chips home. Blackwell GPU production is now underway in Arizona, part of a broader reshoring push that Huang says strengthens national security and cuts reliance on Taiwan. The company’s new BlueField-4 DPU, with 800 Gbps throughput and 6x the compute, will offload networking and security work, freeing GPUs for pure AI acceleration.

Quantum computing took center stage with NVQLink, a high-speed, low-latency interconnect bridging GPUs with quantum processors. Designed alongside 17 quantum hardware startups and nine U.S. national labs, NVQLink enables hybrid quantum-classical computation and real-time error correction. It’s the first real attempt to make quantum practical, using the CUDA-Q stack as the bridge between QPUs and AI supercomputers. Jensen Huang described NVQLink as the “Rosetta Stone” connecting classical and quantum systems.

Nvidia is also moving into telecom, investing $1 billion in Nokia and rolling out its ARC platform to power 5G and 6G networks with AI-driven efficiency. Strategically, Nvidia emphasizes global adoption of U.S. technology, with $500 billion in advanced chip bookings and plans to re-enter the Chinese market. Read more.

Microsoft turns office workers into app builders 

Microsoft’s new Microsoft 365 Copilot Frontier Program drops a full suite of self-building AI tools designed to make software creation as easy as typing a sentence. App Builder can spin up complete business apps on Azure while Workflows automates daily grind tasks like emails, approvals, and scheduling across Teams and Outlook.

A lighter Copilot Studio lets users prototype their own AI assistants trained on company data before scaling them into multi-agent systems. Everything runs under Microsoft’s enterprise-grade security and a new “agent inventory” dashboard for control freaks and compliance teams alike. It’s a bold push to democratize app creation and automation for hundreds of millions of workers, essentially bringing no-code tools to the heart of Microsoft 365. Read more.

• Google launches Pomelli and Vertex AI Training, bridging marketing and model training

• Adobe Max 2025: all the latest creative tools and AI announcements

• Qualcomm leans on mobile DNA to take on Nvidia in AI data centers

• Claude just turned Excel into a Wall Street co-pilot

• You’ll be able to pay with PayPal in ChatGPT next year

• Meta appoints insider Vishal Shah to key AI role

• The Internet is entering its fourth phase, where AI agents talk to each other and the network itself begins to think

• Type a prompt, hit stream. Odyssey-2 turns your words into interactive video in seconds

• Senators propose banning teens from using AI chatbots

• China’s humanoid robot pulls 3,086-pound car with power and control

• GitHub launches Agent HQ, turning its platform into a command center for managing AI coding agents from OpenAI, Anthropic, and Google

• Cursor 2.0 is coming soon

• Goldman: Small businesses using AI say it has a positive impact — and not because it is replacing workers

• Uber partners with Nvidia to deploy 100,000 robotaxis

• Flowith drops flowithOS beta, an AI-native “agentic OS” that blurs the line between browser, assistant, and automation hub

• Grokipedia is now live 

• 1X’s $20,000 Neo home robot is now available to order, promising household autonomy by 2026

• AI use makes us overestimate our cognitive performance, study reveals

• Is AI ready for the courtroom? New framework tackles the technology's biggest weaknesses

• MiniMax’s new M2 model is open-source, agent-ready, and claims 2x the speed of Claude Sonnet at just 8% of the cost

• IBM’s Granite 4.0 Nano proves powerful AI doesn’t need the cloud

• Intuit’s new AI agents for QuickBooks are built on one hard lesson: trust is everything

• Amazon to lay off 14,000 corporate employees

• Fake receipts made with AI are slipping into corporate expense reports, and finance teams can’t trust their eyes anymore

• Can’t afford a vacation? There’s now an AI app that’ll sell you fake photos of one

• Real estate listings are getting weird. AI is now hallucinating stairs, rooms, and entire homes that don’t exist

• China’s military is quietly building AI-powered weapons with help from Deepseek and Alibaba, Reuters reports

• Workers who embrace AI tend to like their jobs more, according to BCG’s chief AI ethics officer

• LexisNexis CEO says the age of AI law is already here

• JPMorgan offers staff AI chatbot to help write performance reviews

• Artificial Analysis: Generative media hits mass adoption in 2025, with 89% using AI for images

• Pinterest experiments with new AI-powered personalized boards

• CFOs must build AI data ‘audit discipline’

• NYU researchers built an AI that flags shady or unenforceable contract clauses before you sign

📄 From Denoising to Refining: A Corrective Framework for Vision-Language Diffusion Model

📄 WorldGrow: Generating Infinite 3D World  ​​

📄 Readers Prefer Outputs of AI Trained on Copyrighted Books over Expert Human Writers

📄 Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations

📄 Video-As-Prompt: Unified Semantic Control for Video Generation

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!