Good morning. It’s Friday, October 3rd.

On this day in tech history: In 2007, Nvidia released CUDA 1.0 to developers, exposing general-purpose compute on GPUs. At the time it was pitched for scientific computing, but it effectively laid the hardware foundation for the deep learning explosion six years later.

In today’s email:

• OpenAI’s Wild Valuation

• Google’s Jules

• Copilot coming for Gmail and Outlook

• 5 New AI Tools

• Latest AI Research Papers

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

OpenAI is the world's most valuable private company after private stock sale

OpenAI has officially become the world’s most valuable private company, hitting a $500 billion valuation after employees sold $6.6 billion worth of shares in a secondary sale. Because the cash went directly to shareholders, not the company, it doubles as a retention play to keep talent from drifting to competitors like Meta’s resurgent AI lab. The milestone comes just weeks after a $40 billion raise in August and sits alongside massive infrastructure bets.

The Sora app is already a viral hit, pulling 56,000 day-one downloads and climbing to No. 3 on the U.S. App Store despite being invite-only and limited to the U.S. and Canada. It’s outpacing Claude and Copilot and matching xAI’s Grok launch, with only ChatGPT and Gemini posting stronger day-one numbers.

ChatGPT itself is now edging into social territory. Recent Android beta builds reveal profile customization. Photos, display names, descriptions and code references to direct messaging further hint at upcoming capabilities. This could allow users to communicate and collaborate directly, echoing what’s already available inside ChatGPT Teams and Workspace accounts. Read more.

Google rolls out Jules CLI/API, Gemini “nano-banana” model now with 10 aspect ratios

Gemini 2.5 Flash Image, aka ‘Nano-Banana,’ is now in general availability, supporting ten aspect ratios, from cinematic 21:9 to vertical 9:16. Developers can blend multiple images, preserve character fidelity across outputs, and make precise edits using natural language prompts.

Early adopters are already pushing it: Cartwheel combines it with 3D posing tools for accurate character rendering, while Volley generates live game visuals with sub-10-second latency for multi-turn edits. Gemini 2.5 runs via the Gemini API on AI Studio or Vertex AI, with deployment or GitHub export, priced at $0.039 per image and $30 per 1M output tokens for scale.

Meanwhile, Google’s coding agent Jules now works in the terminal with full CLI and API access. Developers can trigger, monitor, and parallelize tasks, integrate with Slack or CI/CD pipelines, and manage files, memory, and environment variables, keeping complex workflows smooth without leaving the console.

Rounding out the week, DeepMind’s Dreamer 4 set a new offline RL milestone: it mined diamonds in Minecraft using 20,000+ pre-recorded actions without touching the live game. Its high-efficiency transformer world model folds reward prediction and policy optimization into a single architecture. On the offline diamond benchmark, it beats OpenAI’s VPT and Gemma 3 using roughly 100x less data, delivering faster, more robust learning entirely in simulation. Read more.

Microsoft unifies agents, attacks biosecurity gaps, and arms Copilot for Gmail and Outlook

Microsoft is formally retiring AutoGen and Semantic Kernel, shifting all new development to its unified Agent Framework. The new SDK consolidates multi-agent orchestration (including Magentic One), OpenAPI integration, cross-runtime messaging, and local testing. Governance is embedded at the core: prompt injection shielding, task adherence checks, and PII detection feed into Azure AI Foundry for deployment, monitoring, and lifecycle controls. OpenTelemetry hooks give enterprises visibility into latency, cost, and real-time agent behavior.

Microsoft 365 Premium, meanwhile, is taking a swing at ChatGPT Plus with a $19.99 plan that folds AI into the full Office suite. Subscribers get desktop Word, Excel, PowerPoint, Outlook, and OneNote, plus Copilot and two agents previously limited to enterprises: Researcher, powered by OpenAI’s Deep Research model, and Excel’s Analyst for turning messy data into structured outputs. The tier also unlocks GPT-4o’s highest rate limits, Vision, image generation, Actions, Podcasts, upcoming Photos Agent, Agent Mode, Office Agent, Defender, and 1TB of storage. Personal, Family, and Copilot Pro users are being upgraded automatically.

In a new study, Microsoft researchers have flagged a new biological “zero-day” risk: AI-designed toxins that slip past current DNA screening. In a digital red-team test, Eric Horvitz’s team digitally redesigned known toxins so their genetic signatures wouldn’t trigger safeguards used by DNA synthesis providers. Vendors were alerted and patches issued, but researchers admit some AI-designed molecules can still slip through, framing biosecurity as a moving target. Critics argue defenses must shift into the AI models themselves, not just screening pipelines.

Copilot is also quietly gaining deeper integration and new personalities. “Copilot Connectors” will let the assistant search, read, and analyze content across Outlook, Gmail, Google Drive, Calendar, and Contacts via user-controlled toggles. Outlook integration is designed as a true inbox and calendar agent, hinting at a coming email assistant. A new chat persona, “Coco,” is in testing with a warmer, lifestyle-oriented tone and may roll out alongside up to eight selectable interaction modes. Read more.

• Anthropic hires new CTO with focus on AI infrastructure

• Perplexity’s Comet goes free while Max users get a ‘background AI assistant’

• Mira Murati’s Thinking Machines Lab launches its first product

• Grok is looking for a human “tutor” on how to make video games 

• Apple shelves Vision headset revamp to prioritize Meta-like AI smart glasses

• Netflix will pay you up to $700k per year—and let you work fully remote—if you can harness AI to make employees more productive

• Harvard researchers develop first ever continuously operating quantum computer

• Zhipu AI’s new GLM-4.6 tops Deepseek, trails Sonnet in code tasks 

• Watch: ‘Global Objects’ project uses Azure and NVIDIA to build the world’s largest photorealistic 3D asset library

• Concrete “battery” developed at MIT now packs 10 times the power

• Gaia sets a new benchmark for AI calls with sub-second latency and human-level flow

• Watch: $80,000 humanoid robot brutalized by YouTuber in shocking test

• Princeton AI restores missing fusion data to improve reactor control

• Music's biggest labels are weeks away from major AI licensing deals

• Salesforce launches AI 'trust layer' to tackle enterprise deployment failures plaguing 80% of projects

• Nuclear fusion, the ‘holy grail’ of power, was always 30 years away—now it’s a matter of when, not if, fusion comes online to power AI

• Hume AI released Octave TTS 2 model with multi-language support and EVI 4 mini, designed for live translations

• Pathways’ new AI model “Dragon Hatchling” is inspired by how brains learn

• Ex-OpenAI researcher dissects one of ChatGPT's delusional spirals 

• 'Western Qwen': IBM wows with Granite 4 LLM launch and hybrid Mamba/Transformer architecture

• Intel in early talks to add AMD as foundry customer

• Meta plans to sell targeted ads based on data in your AI chats

• Anthropic integrates Claude with Slack, enabling direct app access and smarter message referencing

• Watch: China’s new Wuji Hand delivers 20 joints, precision grip, and 20kg lift in a 600g frame at $5.5K

• Google opens applications for global AI film award with $1 million prize for AI film creators

• Watch: Leju Robotics opens 10,000m² training hub to close AI’s reality gap

• Engineers create first artificial neurons that could directly communicate with living cells 

• Citi begins retraining 175,000 employees in working with AI: 'great prompting versus basic prompting to generate impactful results'

• Amazon says its AI will let NBA fans track brand-new stats

• Sony adds audio sharing to flagship headphones with new firmware update for Android users

• 'Embodied' AI in virtual reality improves programming student confidence

• AI model can manipulate time to make better predictions in a wide range of fields

📄 MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Use

📄 GEM: A Gym for Agentic LLMs ​​

📄 Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play

📄 DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search  

📄 DA2: Depth Anything in Any Direction 

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!