In partnership with
Reading time: 6 minutes
🗞️ In this edition
Google's Own Researchers Are Quitting Over The Pentagon
Sponsored: Gladly
Anthropic's Agent Contacts You First
Google Will Sell OpenAI's Models
🔌 MCP CORNER Three servers worth wiring up this week.
⚡ YOUR MOVE For people building the life they want, now.
📢 What else is happening
Two of the people building Google's models decided they could not stay. Anthropic shipped an assistant that opens the conversation itself. And Google agreed to start selling its biggest rival's models.
Keep reading.
Source: Ynetnews
The contract lets the Pentagon use it for any lawful purpose. That was the breaking point.
WHAT HAPPENED
Two researchers have resigned from Google's AI division over a classified agreement that lets the Pentagon use the company's models for any lawful government purpose. One of them, a research engineer in London, said she was ashamed of how far Google has moved into military work, and described colleagues who were afraid to speak because they assumed somebody was watching. The other said the company has no real oversight of what happens to a model once it has been handed over. More staff are said to be preparing to leave.
WHY IT MATTERS
Every big lab has published a policy saying what its models may not be used for, and this is what one of those policies looks like when the customer is classified. The people best placed to say whether a system is reliable enough to inform a military decision are the people who built it, and the ones objecting are leaving rather than being answered. That is a safety problem wearing a staffing problem's clothes.
OUR TAKE
The obvious read is that this is a few people with a political objection to defence work. We think the giveaway is that the sharper complaint is about oversight rather than about war, because that one is checkable and the other is not. An engineer saying nobody knows what happens to the model after delivery is describing a hole in the company's own controls, not a difference of opinion about foreign policy.
Watch whether Google says anything in public about how classified deployments get reviewed. If it answers the oversight question, the resignations were about ethics and the controls were already fine. If it answers only the ethics question, the engineers were right and the hole is still open.
From Our Sponsor:
Zero Results, Zero Sales
They came to your site ready to buy. They typed exactly what they wanted. And your search bar ghosted them.
Traditional site search relies on rigid, exact keyword matching. When a shopper uses natural language or makes a simple typo, they hit a “0 Results Found” wall and leave for a competitor.
Stop letting search bars ghost your best customers. Gladly understands shopper intent and turns dead ends into recommendations that keep shoppers buying.
Source: MacRumors
It checks your apps while you are asleep.
Anthropic has released an assistant that starts the conversation instead of waiting to be opened. Claude Duet watches the apps you connect to it, sends a brief each morning, and can go away and finish a job in its own cloud browser, checking with you before it sends a message or spends money. It is a waitlist preview, limited for now to paying subscribers in the United States.
When you do click a story in here, what makes you?
Source: SiliconANGLE
The rivalry lasted right up until the invoicing made sense.
Google has agreed to sell OpenAI's newest models inside its own cloud, next to the Gemini models it builds itself. Three of them arrive in Google's model catalogue within weeks, and companies will be able to pay for them out of cloud commitments they have already signed. The two have spent the past year being described as each other's most dangerous competitor.
🔌 MCP CORNER
THREE SERVERS WORTH WIRING UP THIS WEEK.
An MCP server plugs your assistant into another app, so it stops describing the work and starts doing it.
Railway MCP Server: Ships, checks and rolls back your own services from a chat window instead of a dashboard.
Docker Hub MCP: Lets your assistant find the right container image and tell you what is inside before you pull it.
Airtable MCP Server: Opens your own bases to your assistant, so your records stop needing copy and paste.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🧊 Mistral has put its biggest model into public preview and says the weights land before the end of this month, trained on European machines in a deepening tie-up with Nvidia. If you have been waiting for a frontier scale model you can actually hold in your hands, this is the one to watch. → Try the preview
🎓 Perplexity and Nvidia trained one model to play teacher and student at the same time, and matched the standard reinforcement learning method on roughly a sixth of the training tokens. The paper is up, the code is not out yet, and the method is simple enough to copy. → See how they did it
🧾 Somebody has started selling finance roles rather than prompts. Sixty one of them, each with a job description, approved procedures, limits on what it may do without a human, and a journal an auditor can read line by line. → Read the role list
🔑 The standards for letting one company's agent sign in to another company's systems are unfinished, and the main draft expires next month. There is one thing worth doing now anyway, which is giving every agent you run its own short lived identity and an audit trail. → Lock down your agents
AI Found Two Critical Flaws In PHP – Researchers using frontier AI models found two flaws in the language behind most of the web that let an attacker run code without logging in, and Amazon's decoy servers logged somebody trying one of them four days before the fixes were announced.
A Youth Group Handed Congress A Deadline – Design It For Us has published a 2027 platform asking for independent testing of the largest models, liability when a product injures someone, and the power for government to block a dangerous release.
OnStar Is Turning Into An Assistant – General Motors is rebuilding its in-car service as a voice assistant that works the controls, plans charging stops and explains warning lights, arriving first on one 2027 Cadillac over the air.
Base44: Describe the app you need and get a working one back, without writing any code.
Ideogram: Generates images where the words inside the picture actually come out spelled right.
Elicit: Ask a research question and get real papers back with the findings already pulled out.
Reclaim: Guards your focus time on the calendar instead of handing it to whoever books first.
Three stories about who is still allowed to say no. The engineers who built Google's models could not get an answer about where those models end up, so they left. Anthropic built an assistant that acts first and checks afterwards. And Google decided a customer who wants a rival's model should get it rather than go somewhere else. In each one the person with the clearest view of the risk is not the person making the call.
Work out who in your own setup can still say no.
Interested in featuring your services with us? Email us at [email protected]
