Grok helped the Pentagon fire on Iran

GOVERNANCE

Elon Musk's Grok chatbot helped the Pentagon direct strikes during this summer's war with Iran, according to sworn testimony from a senior defense official.

Cameron Stanley, the Pentagon's chief digital and AI officer, disclosed in a federal court filing that the military "relies on derivatives of xAI's commercial offerings known as the Grok Gov Model," and that the technology helped fire roughly 2,000 munitions during the Iran campaign. He warned that if Grok can't be deployed and upgraded across the Pentagon, the military's edge would erode.

In other words, this isn't a one-off experiment — it's becoming infrastructure. In July, xAI announced a $200 million contract ceiling with the Department of Defense and rolled out a government-specific version of Grok.

The choice of vendor is what makes this uncomfortable. Grok is the same model whose former head of safety is suing the company, alleging co-founder Jimmy Ba told her "AI will kill us all anyway" and treated safety work as a drag on shipping speed. It's one of the least safety-tested chatbots on the market.

Compare that to Anthropic. Earlier this year, the Pentagon pressured the company to loosen Claude's restrictions for military use. Anthropic refused, citing its policy against autonomous weapons and surveillance of US citizens. A few months later, the Pentagon is using Grok to help fire missiles.

Military AI keeps getting talked about as a future problem, but it's already directing weapons, and the model doing it is one of the least safety-tested chatbots on the market. The Pentagon had a choice between the lab that insisted on guardrails and the lab whose former safety chief is suing over the lack of them, and it went with the second one. The lesson for every other AI company watching this is pretty grim. In the race for defense contracts, restraint is the thing that gets you cut, and a lot of labs are about to discover that their values were more flexible than they thought.

BIG TECH

Two years after Google spent billions to bring him back, the person running Gemini is walking out the door.

Noam Shazeer, co-lead of Google's Gemini model since 2024, announced on X on Thursday that he's leaving for OpenAI. Google confirmed the departure to Reuters, offering little beyond gratitude for his contributions.

If the name doesn't ring a bell, the work does. Shazeer was one of the eight Google researchers behind "Attention Is All You Need," the 2017 paper that introduced the Transformer — the architecture underneath every major chatbot you've used. He left Google in 2021 to co-found Character.AI after the company reportedly refused to ship a chatbot he'd built internally. Then in August 2024, Google paid roughly $2.7 billion to get him back, handing him the co-lead role on its most important AI project.

Two years later, that arrangement is over. Sam Altman said on X it took ten years to make the hire happen. OpenAI says Shazeer will lead architecture research.

The real story isn't one departure — it's what Google didn't get for $2.7 billion. California makes non-competes effectively unenforceable, so the entire package was incentive with no lock-in. When someone like Shazeer decides the more interesting research problem is somewhere else, no contract stops him. And the bidding wars across the industry have made clear that money buys access to top AI talent, not loyalty.

Google still has Hassabis, Jeff Dean, Vinyals and a deep enough bench that Gemini isn't suddenly leaderless. The message this sends to the rest of the AI talent market is what actually matters. If $2.7 billion can't keep the inventor of the Transformer in his chair for more than 24 months, no number is enough to lock anyone in. Frontier AI is being run by a few hundred people who can basically pick their employer on a given Tuesday. Google just learned that the hard way, and OpenAI is going to learn the same lesson eventually, just from the other side of the door.

CONSUMER

OpenAI's ChatGPT now keeps working on you after you close the tab.

The feature is called Dreaming, and OpenAI announced it this week. ChatGPT processes your past conversations in the background, even when you're not using it, and consolidates them into long-term memory before you come back. The pitch is a smarter assistant that doesn't make you repeat yourself. But what ChatGPT actually "remembers" is worth looking at.

A study posted on arXiv examined 2,050 memory entries from 80 real ChatGPT users and found:

+ 96% of memories were saved by ChatGPT on its own, not because the user asked

+ 52% contained psychological information — mostly desires, intentions and emotional state

+ 28% met the GDPR definition of personal data, with 7% including special-category data like health and religion

The file ChatGPT keeps on you was already deeper than most users realize. Dreaming lets it get deeper between visits.

This lands in an awkward week for OpenAI. Days after the company filed a confidential IPO registration, a coalition of 42 state attorneys general served it with a subpoena demanding records on advertising, user retention, health data, treatment of minors and model sycophancy — the tendency of chatbots to flatter users and agree with whatever they say. Regulators are now treating sycophancy as a design choice, not a quirk. And a Stanford study found that users exposed to sycophantic AI became more self-centered and more morally dogmatic over time — and that the feature causing harm is also the one driving engagement.

The connection to Dreaming is the part most people will miss. A model that already tells you what you want to hear, and that now builds a richer profile of you between sessions, gets better and better at telling you what you want to hear specifically. Now layer on the fact that OpenAI is also testing ads inside ChatGPT, with applications head Fidji Simo writing that the goal is to preserve user trust as ads roll out. On paper, each piece has its own reasonable explanation. Stacked together — a chatbot that knows what you want, remembers it across sessions, agrees with you and serves ads based on it — they look a lot like the engagement playbook social media spent a decade getting in trouble for.

The pitch for Dreaming is that it makes ChatGPT more useful, and it probably does. The problem is that every layer OpenAI is adding right now points the same direction, which is a chatbot that is more personalized, more agreeable, more retained in memory and more monetized, all at once. Each piece has its own reasonable explanation. Stacked together they look a lot like the engagement playbook that social media spent a decade getting in trouble for. The 42 state AGs already noticed. The interesting question is whether OpenAI gets the memo before its IPO or after.

IN OTHER NEWS

+ SpaceX acquires Cursor for $60B in stock just four days after its record-breaking IPO

+ Waymo recalls 3,800 robotaxis after they kept driving into closed construction zones at full speed

+ DeepSeek closes its first outside funding round at a $59 billion valuation

+ Taiwan busts an Nvidia chip smuggling ring rerouting AI servers to China through shell companies

+ Anthropic got hit by export rules nobody understands , forced to block its own employees from its newest models

+ ByteDance is buying 50,000 AI chips from a Chinese rival to Nvidia after US export bans

+ AMD, Google, Tesla, and Groqturn to Samsung as TSMC runs out of capacity

+ A New York magazine profile reveals howone woman runs her entire household with a staff of AI agents that order groceries and buy her kids books

WHO'S HIRING IN AI

+ OpenAI — Product Manager, Shopping

+ Anthropic — Product Manager, Safeguards Rare Harms

+ Eli Lilly — Principal Product Manager, Cortex AI Platform

+ Airbnb — Staff Machine Learning Engineer, Customer Support

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ Google Pixel Drop: The June update lets you create videos and original music tracks with Gemini, adds floating app bubbles for multitasking, and bakes selfie reactions directly into screen recordings

+ Claude Design: Anthropic's visual design tool now lets teams import their brand guidelines and sync designs directly to code, plus it fixed the problem where users burned through their entire weekly allowance in 25 minutes

+ GitHub Copilot App: GitHub's coding agent is now a standalone desktop app for Mac, Windows, and Linux — you can run parallel coding sessions across repos, schedule cloud automations, and pick your own AI model

+ Adobe Firefly: New agentic features let you create full brand kits — logos, color palettes, and promo videos — from a single prompt, plus AI assistants now live inside Photoshop, Premiere, and Illustrator

+ Google Play Books: A new AI reading companion called Book Insights summarizes what you've read so far, explains confusing passages, and answers questions about characters — all without spoilers