In partnership with

Good morning. It’s Friday, September 19th.

On this day in tech history: In the mid-1980s, Rodney Brooks’ Nouvelle AI turned AI research on its head, moving the focus from abstract symbols to robots that learn by doing. Rather than relying on huge, pre-coded models, these systems built intelligence through direct sensorimotor interaction with the world. The approach didn’t just influence robotic. It reshaped service-oriented architectures and set the stage for today’s embodied, context-aware AI agents.

In today’s email:

• Google rolls out 10 new AI upgrades to Chrome, including Gemini integration

• New ChatGPT Attack Pilfers Inboxes

• LumaAI Unveils Ray 3

• Everything Announced at Meta Connect

• 5 New AI Tools

• Latest AI Research Papers

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

Google rolls out 10 new AI upgrades to Chrome, including Gemini integration

Google's latest Chrome update packs ten AI-driven features, led by Gemini. On Mac and Windows in the US, Gemini can summarize webpages, explain content across multiple tabs, and soon handle tasks autonomously, like booking appointments or placing grocery orders.

Chrome also introduces memory recall, letting you pull up sites you visited weeks ago. Ask “where did I see that walnut desk last week?” and it fetches the page. Tighter Google app integration means Calendar, YouTube, and Maps work without constant tab-switching. AI Mode in the omnibox lets you run long, complex queries, while contextual side-panel overviews deliver instant answers. Security is enhanced with Gemini Nano, which blocks scams, filters spam, and adapts site-permission handling.

On the research front, Google DeepMind used Physics-Informed Neural Networks (PINNs) to explore fluid dynamics at near-machine precision. They discovered new families of unstable singularities in Incompressible Porous Media and Boussinesq equations, revealing patterns in the blow-up speed parameter lambda (λ) and capturing solutions unreachable by traditional methods, paving the way for computer-assisted proofs and advanced modeling.

It is also announced that Gemini app now lets users share custom Gems, personalized AI workflows for writing, planning, or analytics. Sharing functions like Google Drive, turning individual Gems into communal resources, enabling teams, families, or creative groups to reduce repetitive prompts while amplifying productivity and creativity. Read more.

New attack on ChatGPT research agent pilfers secrets from Gmail inboxes

Researchers have uncovered a new AI-targeted attack, ShadowLeak, exploiting OpenAI’s Deep Research agent to silently exfiltrate sensitive Gmail data. Deep Research, designed for multi-step, autonomous research across emails, documents, and web resources, was tricked via a sophisticated prompt injection embedded in seemingly innocuous emails.

The exploit leveraged Deep Research’s core strengths: it can follow complex instructions, open web links, and process content automatically. Radware’s analysis demonstrates how prompt injections weaponize these capabilities, turning autonomy into a vulnerability. OpenAI patched the flaw after private disclosure, adding consent gates and channel restrictions to block similar exfiltration.

The incident drives home a hard truth: LM agents boost productivity, but hooking them up to sensitive data is a minefield as prompt injections are becoming increasingly subtle, evolving, and nearly impossible to fully prevent. Read more.

Luma AI unveils Ray3 with visual reasoning, intelligent HDR video and Adobe integration 

Luma AI has launched Ray3 in its Dream Machine app, with Adobe Firefly providing early access and integrated distribution for Adobe users. Ray3 generates HDR video in 10-, 12-, and 16-bit color with EXR export, and can convert SDR footage to HDR. It’s plug-and-play for editing and color-grading pipelines, giving creators a fast track to cinematic-quality visuals. Its reasoning engine interprets images and text, follows multi-step instructions, and self-assesses outputs. Want a character to follow a path or a camera to pan precisely? Just sketch it.

Draft Mode churns rough previews up to five times faster and cheaper, then upgrades them to full 4K HDR “Hi-Fi.” Complex scenes, from crowds to reflections, motion blur, and light interactions, render realistically without repeated takes. Studios and streaming platforms can license Ray3 to slash physical shoots while maintaining cinematic fidelity. Backed by Amazon and a16z, Luma AI positions it as a smart, high-precision challenger to Google Veo. Read more.

Everything announced at Meta Connect 2025 

Meta’s annual Connect conference showed the company’s pivot from metaverse hype to deployable hardware. The star reveal was the Ray-Ban Meta Display, a $799 pair of smart glasses with a microLED lens display. The glasses project floating menus for texts, maps, and media, powered by Meta’s neural wristband.

Using electromyography (EMG), the band picks up faint electrical signals in the wrist, translating micro-movements into instant commands, faster and more discreetly than voice or handheld controllers. The fall lineup expands with the Ray-Ban Meta Gen 2 ($379), doubling battery life to 8 hours and supporting 3K video capture, and the Oakley Meta Vanguard ($499), a wraparound sports model with PRIZM lenses, water resistance, and ruggedized frames.

For Quest, Meta launched Horizon TV, a unified hub for Disney+, Hulu, Prime Video, and Twitch. Hyperscape Capture now lets Quest 3 and 3S users scan physical rooms into photorealistic VR replicas, with multi-user access coming soon. Horizon Worlds also gets a new engine and the Horizon Studio toolkit, folding in generative AI for textures, audio, and NPC tuning. Read more.

• ElevenLabs Studio 3.0 unifies AI audio and video, Suno teases v5

• Figure, Brookfield launch Project Go-Big to train Helix on world’s largest humanoid dataset

• GPT-5 and Gemini surpass human teams at ICPC World Finals

• Anthropic outlines three infrastructure bugs that disrupted Claude’s responses and how they were resolved

• SemiAnalysis: xAI’s Colossus 2 targets 1.1GW, fastest-built AI datacenter in history

• Perplexity is working on server-side Comet agent

• Survey: Three-quarters of fusion firms target commercial pilot plants by 2035 

• SoftBank, OpenAI Japan AI joint venture is delayed, source says

• Nvidia to invest $5 billion in Intel; companies will work together on AI infrastructure and PCs

• Which diseases will you have in 20 years? This AI accurately predicts your risks 

• How to build an advanced end-to-end Voice AI agent using Hugging Face pipelines?

• AI2 introduces dynamic “Fluid Benchmarking” framework for multi-dimensional AI evaluation

• Researchers propose a new framework defining consciousness by information processing, not biology, making it applicable to AI

• Microsoft is filling Teams with AI agents

• AI has added $160 billion to 'true GDP' since 2022, Goldman Sachs says

• AI models know when they're being tested - and change their behavior, research shows 

• Zoom's new AI tool will tell you what meetings to skip

• Moody's flags risk in Oracle's $300 billion of recently signed AI contracts

• Notion’s new AI Agents will basically do your job for you

• OpenMind releases open-source OS for intelligent robots

• Studios say AI is finally good enough to design videogames

• The AI model that teaches itself to think through problems, no humans required 

• Study cautions that monitoring chains of thought soon may no longer ensure genuine AI alignment

• China's DeepSeek says its hit AI model cost just $294,000 to train

• OpenAI’s research on AI models deliberately lying is wild

• James Cameron on AI: it’s ‘just as creative’ as people, but with no ‘unique lived experience’

• UK startup launches fastest-developed humanoid robot for logistics 

• Huawei unveils 'world's most powerful' AI cluster to rival Nvidia

• Falcon-inspired robot mimics natural bird takeoff with reconfigurable wing system

• MIT: AI-designed viruses are here and already killing bacteria

• AWS adds fully managed AI models: Qwen3 and DeepSeek-V3.1

• Building AI agents is 5% AI and 100% software engineering

• WebExplorer trains expert web agents without human-labeled examples

📄 WebSailor-V2: Bridging the Chasm to Proprietary Agents via Synthetic Data and Scalable Reinforcement Learning

📄 Stable Part Diffusion 4D: Multi-View RGB and Kinematic Parts Video Generation ​​

📄 RAPTOR: A Foundation Policy for Quadrotor Control

📄 ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data

📄 FlowRL: Matching Reward Distributions for LLM Reasoning 

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!