In partnership with

Good morning. It’s Wednesday, October 15th.

On this day in tech history: In 1979, a group at MIT’s Architecture Machine Group logged a test of their “Put-That-There” multimodal interface using a new gesture parser. The notes from that week mention the system randomly triggering on coughs, an early lesson in multimodal false positives.

In today’s email:

• Google’s Newest AI Releases

• OpenAI To Lower Guardrails

• NVIDIA Ships Tiny Supercomputer

• 5 New AI Tools

• Latest AI Research Papers

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

Google Ships Image Gen in Search, Agents in Enterprise and Real-Time AI in Meet

Nano Banana, the Gemini 2.5 Flash–based model behind 5B+ image edits, is now embedded across Search, Lens, NotebookLM and soon Photos. Lens on iOS and Android gets a new Create mode for real-time image generation and transformations.

NotebookLM now uses Nano Banana to auto-illustrate video overviews in six styles and can generate Brief clips that compress dense documents into narrated, visual capsules.

Search and Discover are moving into real-time surfacing. Discover now pushes expandable previews of trending topics linking straight to articles and a “What’s new” feed for player and team lookups, mixing live stats with media pulls. Gmail’s new Help me schedule parses draft intent and calendar availability to surface suggested slots inline and auto-issue invites once accepted. One-to-one only for now.

AI Studio now has a live token/rate dashboard and an unreleased “I’m Feeling Lucky” generator that assembles random prompts using Gemini modules like Nano Banana. Model selector hints at variant unlocks.

And Google Meet now has virtual makeup that actually tracks your face, powered by real-time face landmark models, and stays put even when you move. It’s off by default but rolling out on web and mobile, finally catching up to Teams and Zoom. Read more.

OpenAI makes ChatGPT human, shoppable, and adult-ready

The company has launched its Expert Council on Well-Being and AI, a panel of eight specialists in psychiatry, psychology, and human-computer interaction, to advise on safe AI usage. Council insights are already shaping ChatGPT and Sora, powering parental controls and an age-prediction system that automatically applies teen-appropriate settings and alerts parents if a child shows signs of distress. This comes amid an FTC inquiry and lawsuits questioning the safety of AI interactions for minors.

For adult users, a major December update will let ChatGPT feel more human and expressive. Verified adults can engage in erotic conversations, customize tone, personality, and style, and use emoji-enhanced responses. OpenAI has deployed automatic moderation via gpt-5-chat-safety, switching models when users enter sensitive emotional or mental-health topics, while restoring the friendlier, more personable traits from GPT-4o.

A new Walmart partnership lets users browse and buy products, including Sam’s Club items, directly in-chat via OpenAI’s Instant Checkout. This replaces clunky search bars with a more interactive, multimedia shopping experience and positions Walmart to compete with Amazon’s AI tools. OpenAI’s latest streak of massive deals show Sam Altman is selling a vision of a world-changing product with financial engineering to fuel the company’s $1 trillion project ambition. Read more.

Nvidia’s pocket-sized AI supercomputer arrives October 15 

Nvidia’s ultra-compact ‘personal AI supercomputer’ lands today. DGX Spark, starting at $3,999, squeezes a petaflop of FP4 performance into a 2.6-pound chassis. At its heart, the GB10 Grace Blackwell Superchip fuses a 20-core Arm Grace CPU with a Blackwell GPU, RTX 5070 class CUDA cores, and fifth-generation Tensor Cores.

The system packs 128GB of unified LPDDR5x memory, 4TB of NVMe storage, NVLink-C2C delivering 5× the bandwidth of PCIe Gen 5, and an Ubuntu-based DGX OS preloaded with CUDA libraries, NIM microservices, frameworks, and pretrained models.

DGX Spark handles inference on models up to 200B parameters and local fine-tuning on models up to 70B. OEMs including Dell, HP, Lenovo, Acer, Asus, Gigabyte, and MSI will ship branded desktops, while a larger GB300-based DGX Station is coming. Nvidia CEO Jensen Huang personally delivered the first unit to Elon Musk at SpaceX, right next to a Starship pad.

Nvidia also unveiled its “gigawatt AI factories” with the Vera Rubin NVL144 rack server. Liquid-cooled and open-standard, it scales modularly with expansion bays, 800V DC power, and 45°C liquid-cooled busbars. NVL144 supports up to 576 upcoming Vera Rubin GPUs, cableless NVLink-C2C networking, a central PCB midplane, and ConnectX-9 800GB/s links, while Meta and Oracle adopt Spectrum-X Ethernet for high-speed AI training. Together, these launches bridge desktop experimentation and enterprise-scale AI workloads. Read more.

• Anduril’s EagleEye puts mission command and AI directly into the warfighter’s helmet

• World's first portable nuclear reactor factory planned at Manhattan Project site 

• 𝕏 to release Grok-powered algorithm update this week with open weights

• China producing 'world-first' quantum radars to track US stealth jets like F-22 

• Anthropic's Jack Clark compares AI breakthroughs to hammers that suddenly become self-aware

• Oracle adds AI capabilities to core database and launches a lakehouse platform

• "Pixnapping" attack exposes 2FA codes on Android; Google's September patch falls short

• Andrej Karpathy launches nanochat, a minimalist all-in-one LLM framework 

• This digital hand enables hands-free virtual reality 

• Visa just launched a protocol to secure the AI shopping boom — here’s what it means for merchants 

• Mozilla's Firefox adds Perplexity's AI answer engine as a new search option 

• Salesforce introduces new OpenAI, Anthropic integrations

• Goldman tells staff it will cut more jobs as AI saves costs 

• Naked mole rats live for decades — genetic tweaks reveal insights into ageing 

• Why replacing the human touch with AI is a big mistake for companies

• Robot 'backpack' drone launches, drives and flies to tackle emergencies

• Watch: Neuralink co-founder presented a new theory of consciousness last week in Tokyo

• New open-source tool helps engineers create gradient 3D-printed designs

• Apple expected to introduce M5 iPad Pro, base M5 MacBook Pro this week

• Meta and Oracle embrace NVIDIA’s Spectrum-X: Ethernet enters the mega AI factory era 

• Next-generation memory: Tungsten-based SOT-MRAM achieves nanosecond switching and low-power data storage

• This new AI technique creates ‘digital twin’ consumers, and it could kill the traditional survey industry

• Watch: Unitree G1 Kungfu Kid V6.0

• Slack is turning Slackbot into an AI assistant

• Cerebras aims to deploy AI infrastructure for massive Stargate UAE data centre hub

• Self-improving language models are becoming reality with MIT's updated SEAL technique

• California becomes first state to regulate AI companion chatbots

• CEO says she’d welcome an AI-bot board member: ‘If you don’t have an AI agent in every meeting, you’re missing out’

📄 StreamingVLM: Real-Time Understanding for Infinite Video Streams

📄 Instant4D: 4D Gaussian Splatting in Minutes ​​

📄 Thinking with Camera: A Unified Multimodal Model for Camera-Centric Understanding and Generation

📄 D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI  

📄 QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs

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!