Welcome back! Nvidia launched Cosmos 3 Edge, a model that helps robots see and move, and got 10 big Japanese companies on board. Publishers sued Google, saying Gemini was trained on their books, with an internal Google memo reportedly warning of up to $100 billion in fines. OpenAI's $230 keypad is almost sold out, and Waze added Gemini voice search.
In today's Generative AI Newsletter:
Nvidia: What can Cosmos 3 Edge actually do?
Courts: Why is Google being sued over Gemini?
OpenAI: How to keep an eye all your AI agents?
Waze: Who can now boss Waze around by voice?
On July 15 in Tokyo, CEO Jensen Huang unveiled Cosmos 3 Edge, a model built for robots and vision agents. It is a world model, so it learns to perceive and move through physical space in real time, from more kinds of input than a language model.
It runs on-device, on Nvidia's Jetson Thor boards, so a robot can reason without a round trip to the cloud. It joins the Cosmos 3 family of open world models.
Huang also got Japan behind it. Ten big manufacturers, including Fujitsu, Hitachi, Sony and Yaskawa, plan to build on Cosmos. The next day, Nvidia and a company called Noetra went bigger, announcing a factory they will build with 27,500 Rubin GPUs to train open models for robots and digital twins.
This is Nvidia widening its moat past chips. If robots, factories and cars standardize on Cosmos, Nvidia owns the model layer of the physical world too.
Huang called the physical world "the next frontier of AI." Japan, where factories outnumber the workers to run them, is the proving ground.
With hundreds of investors on board, GenAI Works is building the go-to-market layer of the AI economy, the distribution every AI product needs to reach real buyers.
All eyes are on AI.
Anthropic filed to go public, SpaceX paid $60B for Cursor and OpenAI raised more than most countries spend on defense. That's the market power of AI.
As for GenAI Works, the potential is evident in the numbers:
$2.7 million in EOI, money already in the funnel.
2.5x revenue growth last year while running go-to-market for Nvidia, Oracle, Google, IBM and 300+ AI brands.
A 14M+ AI community powering ToneUp, our distribution platform.
Special highlight from our network
Open Claude. Ask a question. Close the tab. That's how 95% of professionals use it. Which is fine, until you realize the other 5% are quietly building a very different kind of workday.
The gap isn't about who's smarter with prompts. It's about who set up Skills so Claude has persistent expertise. Who wired up Connectors so it can pull from Notion, files, and their calendar. Who learned Cowork for full document workflows. Who started vibe coding tools they used to pay engineers for.
Outskill's 2-Day Claude AI Mastery Workshop is a live, end-to-end walkthrough of all of it, condensed from 800+ hours of research into a focused 16-hour curriculum. Plus 10+ AI tools and workflows that pair with Claude.
Saturday and Sunday, 10 AM to 7 PM EST. Free for the next 48 hours.
An internal Google document reportedly put the risk of training Gemini on books at up to $100 billion in fines.
On July 10, Hachette, Cengage, Elsevier, author Scott Turow and S.C.R.I.B.E. filed a class action in the Southern District of New York, accusing Google of training Gemini on their copyrighted books without permission.
The publishers have a long history with Google. They handed over books so Google Books could show short searchable snippets, not full texts. The suit says Google copied those books, plus titles uploaded to Google Play, to train Gemini without authorization.
It goes further, alleging Google removed or altered copyright information to hide that Gemini had been trained on those books.
The filing points to an internal Google document that allegedly flagged the practice as a serious legal risk, with "$10Bs-$100Bs in potential fines."
Two early California rulings favored AI companies on fair use, but a different New York judge now gets to weigh in. Anthropic already paid $1.5 billion over pirated books, the largest copyright payout on record.
OpenAI put out its first physical product, and the limited run is nearly sold out.
The Codex Micro is a $230 keypad for riding herd on AI coding agents, built with keyboard maker Work Louder. It went on sale July 15 in a limited run.
Inside the small aluminum body:
13 mechanical switches
a touch sensor
a rotary dial
a joystick
The dial turns an agent's reasoning up or down. The joystick fires actions like reviewing a PR or debugging an error.
OpenAI started teasing this hardware in June, and a screenless home device is still rumored to come. The keypad is the cheap test of whether people will buy OpenAI hardware. Early signs say yes.
Waze just wired Gemini into the drive.
Google's navigation app added several AI features: you can now search by talking to Gemini, asking things like where to find open parking or the cheapest gas nearby, and Waze returns a list.
It also learns your habits. Personalized navigation suggests routes from your trip history and a city's traffic. You can switch it off in settings.
A new Motorcycle mode routes around two-wheeler restrictions and flags potholes, speed bumps and narrow bridges. It is live in seven countries so far. A "less chatty" mode trims the voice prompts when you would rather hear your podcast.
This is Google pushing Gemini into a daily habit and aiming Waze straight at Apple Maps. Most of it is live now on Android and iOS, with voice search still in beta.
The reporting trick is the sleeper. Say "the road is closed here" and Waze forwards it to local map editors, turning every driver into a live sensor.
NVIDIA Build lets you try frontier and open models in your browser, then copy the API call to run the same one in your app. There is nothing to install and no GPU to rent.
Try this yourself:
Go to build.nvidia.com and sign in with a free NVIDIA account.
Open Models and pick one, or filter by Physical AI or Agentic AI to match today's news.
Type a prompt in the playground and run it to see the output live.
Open the code panel to grab the API snippet and drop it into your project.
Who it's for: developers who want to test a model before committing, and anyone curious to poke at robotics or agent models without buying hardware. The playground is free to try.
No AI lab scored above a C+ on safety: The Future of Life Institute's Summer 2026 index, released July 7, gave Anthropic the top grade at C+, with OpenAI and Google DeepMind at C, Meta at D+ and xAI, DeepSeek and Mistral flat failing.
Anthropic lines up an autumn IPO: Anthropic is reportedly targeting a fall listing to reach the public markets before OpenAI and DeepSeek, on roughly $47 billion in annualized revenue.
TSMC's June sales jumped 67.9% on AI demand: The world's largest chip foundry reported June revenue up 67.9% year over year, and analysts say its N3 node is fully booked, which keeps AI compute scarce and pricey.
Google keeps Gemini 3.5 Pro in the lab: Google held back Gemini 3.5 Pro after testing showed newer builds scored worse than older checkpoints, leaving its July 17 target in doubt.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
