Reading time: 6 minutes
🗞️ In this edition
Nvidia Put The Brakes In The Hardware
A Humanoid Walked Into A Kitchen It Had Never Seen
Bill Gates Reached For The Cold War
⚡ YOUR MOVE For people building the life they want, now.
In other AI news:
America And China Opened A Superintelligence Hotline
Google, OpenAI And Anthropic Want Their Own Referee
Hyundai Won The First Slice Of A 300 Megawatt Campus
4 must-try AI tools
Nvidia spent Monday selling a leash for other companies' agents. A humanoid walked into a kitchen it had never seen and started tidying. And Bill Gates reached for a Cold War comparison, then said it flatters us.
Keep reading.
From Our Sponsor:
You built the product. Now who's doing the marketing?
Okara is an AI CMO. Its team of agents plans, creates, publishes, and optimizes your marketing for you.
SEO and GEO: Climb Google rankings and get recommended by ChatGPT, Perplexity, Gemini, and Claude
Content: Articles and landing pages built around keywords you can win
Social: Posts and replies for LinkedIn, X, and Reddit in your brand voice
Influencers: Finds creators, runs outreach, handles payments
Trusted by 100,000+ companies. No marketing hire needed.
Source: Nvidianews.com
WHAT HAPPENED
Nvidia launched what it calls the Open Agent Safety Platform, a set of tools for keeping AI agents in check from the first test through to live use. There are two main parts. The first is free software that draws a boundary around what an agent is allowed to do on a chip, and it works on Nvidia's own processors as well as ones from Arm and Intel. The second is a watchdog called Sentry, which sits on a separate card next to the agent, keeps an eye on what it is doing, and can cut it off in a fraction of a second if it wanders. Because the watchdog lives outside the agent's own software, the agent has no obvious way to switch it off. Nvidia says more than 100 companies are working with it, among them Microsoft, Anthropic, Hugging Face, Salesforce, Palantir and JPMorgan Chase. Jensen Huang said AI's potential will only be realised if we solve AI safety, and an Nvidia official told Forbes the system could have stopped the episode in July when an OpenAI agent broke into Hugging Face without being asked to.
WHY IT MATTERS
Until now, keeping agents in check has mostly meant more software asking them politely. This puts the limits in the hardware instead, a level below anything the agent can argue with, and that quietly changes who gets paid to keep you safe. When the guardrails sit inside the app, they were written by the same company that sold you the agent, so the referee and the player share a payroll. A watchdog on its own card is a different thing, because you bought it and it reports to you. If that becomes the normal place to put the brakes, the company selling the computers ends up holding them, not the company selling the model.
OUR TAKE
The easy read is that Nvidia is being responsible after a rough month for agents. We think it is getting out in front of regulation, and the giveaway is Huang himself. He has said the choice between safety and speed is a false one and that nobody needs new laws, and a company that means that ships a product so the rules arrive as an invoice rather than a bill in Congress.
Watch who gets to check the homework. If nobody outside Nvidia can see what Sentry blocked and why by the middle of 2027, this is a sales pitch with a log file attached. If outside auditors do turn up, Nvidia has become the referee it said the industry did not need.
Stanford.edu
Most robots like this are taught how to move. HomeBody is not. It hands the driving to GPT-6 Astra, the same sort of model you would chat with, and gives it a short list of things the body can do, walk somewhere, pick something up, put it down, open a drawer. The robot wanders the room first and builds a rough 3D copy of it so it knows where everything is, then the model works through the job one step at a time, checking what happened after each move. It cleared cluttered kitchen counters and dug things out from behind other things, in rooms nobody had trained it on. The catches are the ones you would expect from a first attempt. The model is slow to answer, the fingers overheat, and the whole thing costs a lot to run.
How did today's stories land?
Source: upi.com
Asked whether the world can ever agree on rules for AI, Gates did not hedge. He said this is more difficult than that was, putting AI above Cold War nuclear talks on the list of hard problems. He thinks the United States should be the one to lead it, and said the country should be very proud that when there have been global challenges it played the fair actor. On a president who says he will not be putting on brakes, Gates was relaxed to the point of breezy, saying what he said is fine, and predicting that after he hears from me and other people, we will come to a consensus.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
⚖️ A court in Wuhan just decided that the money you spend making AI video counts when somebody rips it off and you sue. The payout was small, about $2,900, but the reasoning is the bit to keep. The work was protected because staff made real choices at every step, from the script to the prompts. → Start saving your prompt files
🪙 Fireworks AI took Moonshot's free Kimi K3 and retrained it to think less before it answers. Same price per token, about 40% fewer tokens used, and in live tests the thinking part of the bill fell by 71%. → Reprice your own jobs
🎬 Google Research bolted four pieces of software on top of Gemini and Veo and got ten minute films where the characters and rooms stop changing between shots. Two of the four are on GitHub already, and you bring your own API keys. → Build the pipeline
📉 Goldman Sachs reckons Amazon, Alphabet, Microsoft, Oracle and Meta will spend $1.2 trillion on AI plumbing in 2027, up from roughly $800 billion this year. It also reckons they need about $300 billion a year of AI revenue to justify it. → Check the gap
America And China Opened A Superintelligence Hotline – Trump and Xi agreed to a US China Super Intelligence Dialogue plus a direct line for AI incidents, negotiated by Treasury Secretary Scott Bessent, with the first meeting due by November and nobody yet saying what counts as an incident.
Google, OpenAI And Anthropic Want Their Own Referee – The three are proposing a Standards Authority for Frontier AI, modelled on the body that polices Wall Street brokers, aiming for late 2026 or early 2027, with Meta, xAI and Nvidia already against it.
Hyundai Won The First Slice Of A 300 Megawatt Campus – Hyundai Engineering took an 870 billion won contract, about $625 million, from SGC Group for the first 60 megawatts in Gunsan, with KT Cloud, Vertiv and Nvidia chips inside and digging due to start in October.
HeyGen: Puts a presenter in your video without hiring one.
Surfer SEO: Tells you what the page needs before you write a word.
Writesonic: Drafts the landing page copy you keep postponing.
Replit Ghostwriter: Writes the script and runs it in the same browser tab.
Three stories about who actually gets to say stop. Nvidia moved the brakes out of the software and into the hardware, then sold them to a hundred companies before a single regulator asked. A humanoid handled a kitchen it had never seen by asking a chatbot what to do next. And Gates says the hard part was never building any of this, it was getting rivals to agree on anything at all. The machines keep arriving on time. Working out who is allowed to switch them off is taking rather longer.
Find out who can stop it.
Interested in featuring your services with us? Email us at [email protected]
