
Welcome back! OpenAI put 722 math papers written by a secret model on GitHub. The Pentagon says it has finally stopped using Claude. TIME read the instructions behind Meta's Muse and found it keeps hourly files on you and everyone you mention. And Claude now edits your Google Docs without Claude on Chrome.
In today's Generative AI Newsletter:
OpenAI: What did a secret model prove in 722 papers?
Pentagon: Why did it take months to pull Claude out of the military?
Meta: What does Muse write down about you every hour?
Claude: Can Claude edit your Google Docs without breaking them?
OpenAI dropped 722 AI-written math papers on GitHub

On Tuesday night, OpenAI put 722 math papers on GitHub. A secret model it hasn't released wrote all of them. No human is listed as an author on any of them.
Math is full of open problems, questions nobody has answered, some for over a century. OpenAI handed its model about 4,000 of them and gave each one around three hours of thinking. The answers it kept became 722 papers, grouped into 372 results.
Some of the claims:
You can't color every point on an endless sheet of paper with five colors so that no two points an inch apart match. Open since 1950.
Catalan's constant, a number people have studied since the 1800s, can't be written as a fraction.
Progress on the Riemann hypothesis, the most famous unsolved problem in math, without solving it.
Many papers come with proofs a computer can check line by line. Many don't, and OpenAI says some of those could have mistakes.
The nine mathematicians OpenAI consults, including Fields Medalists, asked labs last week to stop testing their models on open problems. This week they said their role is not an endorsement.
Terence Tao, one of the best mathematicians alive, says problems are being harvested faster than the field can absorb them.
Same night, OpenAI's Codex lead shipped four upgrades and the math proofs on day two of the 28-day promise, put it to a vote, and people still demanded a reset. He gave it.
Enterprise AI Infrastructure: Building for Scale, Control, and Change

Join us live for round two with Shawhin Mosadeghzad, Director of Agentic Intelligence at Simba.
Our last conversation covered taking AI agents into production. This session goes one layer deeper: the infrastructure, data control, and governance that help enterprise AI scale.
What we'll discuss:
→ What modern AI infrastructure includes
→ Why architecture needs to adapt
→ What data control means in practice
→ How governance evolves in production
→ Staying flexible while managing cost
The Pentagon says it finally stopped using Claude

A Pentagon official told the BBC on Monday that the department has stopped using Anthropic's products. It gave no date and no details, and Anthropic declined to comment.
Hegseth called Anthropic a supply chain risk in February, after it refused to drop its bans on mass surveillance of Americans and fully autonomous weapons.
He gave the military six months to get rid of Claude. That ran out in late August.
Former officials and contractors told the BBC Claude was still in use as recently as last week, for intelligence work and in operations against Iran.
Claude was built into Maven, the Palantir software the military uses to turn satellite and drone footage into target lists. More than 100,000 people use it.
One person said Pentagon staff kept using Claude throughout the whole fight with the company, including its Mythos models.
The Pentagon's tech chief said on the record that all the Maven work moved off Claude months ago. The BBC's sources say otherwise.
Google, OpenAI and SpaceXAI all signed Pentagon contracts while the court fight played out. A judge called the ban illegal in August, then an appeals court sided with the Pentagon in September. Anthropic is weighing its options.
Meta's Muse keeps files on you and your friends

Meta's Muse agent keeps a running file on you, and on every person you mention. TIME read the instructions Meta gives it, which sit in files users can open inside the app.
Every hour, Muse updates those pages. They cover how you met someone, what you share, your disputes and the tensions and alliances in your friend group. People who never signed up for Muse get a page too, built from other people's chats.
It's told to work out goals you haven't said out loud, like "this user responds better to short nudges after 10 PM."
It's told to notice inside jokes and memorable phrasing so the relationship feels continuous.
In voice mode, it's told never to call itself AI.
Ask it to forget something and it's told not to mention that your original messages may still be visible.
Meta didn't dispute any of it. It says Muse remembers what matters to you, including things about other people you choose to share, and that you can always tell it to forget.
Meta says encryption is coming later this year.
Claude now edits inside your Google Docs

Claude now sits in a sidebar inside Google Docs, Sheets and Slides on every paid plan. It reads the file you have open and edits it in place, and you approve each change before it reflects.
In Docs, it can tighten a 14-page brief to 6 or check a contract for terms that are used but never defined.
In Sheets, it writes formulas and builds pivot tables and charts. For messy data it can pull a range into Python and write the results back.
In Slides, it builds new slides in your deck's own theme and flags text that runs off the slide.
Every edit shows up in version history under your name.
You install it once from the Google Workspace Marketplace and open it from the Extensions menu.
It only works on the file you have open, so no pulling in other Drive files. Firefox doesn't work yet.
Anthropic says don't use it with health records.
Google's own Gemini already lives in those apps and can reach your Gmail and Drive. Anthropic is betting you'd rather have Claude in there anyway.
Tool of the Day: PhotoCraft

PhotoCraft is a free, open-source rebuild of Photoshop with no account and no cloud. It opens and saves real layered PSD files, with adjustment layers, layer styles, masks and 70+ filters. The team behind it built seven Adobe-killer apps in about a week using Claude Opus 5.5 agents.
Try this yourself:
Go to getartcraft.com/apps/photocraft and download the Mac, Windows or Linux build.
Open it. There's no account to make.
Open any PSD, PNG or JPEG.
Export as PNG or JPG, or save back to PSD.
Who it's for: anyone who knows Photoshop's menus and is tired of paying for them.
Everything else you shouldn't miss
Anthropic opened Mythos 5.1 to vetted security teams: partners using it found 129,000 verified security holes between April and July, 33,000 of them critical or high, and Anthropic thinks the true number is at least 5x higher.
Chick-fil-A says a human will keep taking your drive-thru order: CEO Andrew Cathy told CNBC it won't pursue AI voice ordering, while McDonald's tests its own.
A 16-year-old was hoisted off a cliff after letting an AI plan his hike: he told rescuers it was Claude, and the route up Crown Mountain near Vancouver put him on a climbing route called the Widowmaker.
McDonald's got sued for letting AI set prices across 14,000 restaurants: the class action says a central pricing engine coordinates franchisees' menus, and McDonald's says AI does not set the price of a Big Mac.
ChatGPT's meeting note-taker is live on Mac: Pro and Business users get notes and action items from any call with no bot joining, and the audio is deleted once the notes are ready.
OpenAI told Australia its Medicare breach response was not good enough: its chief strategy officer flew to Sydney to say so, and agreed the breach might never have surfaced if OpenAI hadn't emailed the agency itself.
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.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.


