Good morning, and happy Sunday.

This is the weekly catch-up edition. The biggest story ran across four mornings: AI took on real science, from math proofs to a map of the whole sky, and each time a person had to check what it produced. Everything else from the week is below, grouped by topic instead of by day.

THE FRONT PAGE

AI Did Big Science Jobs This Week, and People Caught What It Missed

TLDR: In five days, AI helped settle open math problems, find people at risk of Ebola, fill in a third of the sky and screen urgent-care patients, and in every case a person had to check the work.

The Story:

On Monday, Meta said mathematicians using its regular Muse Spark chat app co-wrote six papers, and five answer questions nobody had settled, with people checking every step. On Wednesday, Google said its Earth AI tools helped the World Health Organization find more than 45,500 people at risk of Ebola in Congo in minutes instead of weeks. On Friday, Johns Hopkins astrophysicist Brice Ménard used Claude Science to build a full map of the sky in ultraviolet light, predicting the third no telescope has seen to within about 10% on test patches. Ménard also does research at Anthropic, which published the account, and the map hasn't been peer reviewed. The same day, OpenAI posted 722 math papers from an unreleased model, and only 162 have a main result checked by Lean, a program that verifies proofs. On Saturday, a Lancet study of Google's medical chatbot AMIE found its single top guess matched the final diagnosis 56% of the time, with a doctor watching every chat.

Its Significance:

In two of these, a person found mistakes the AI missed. Faint smudges from atmospheric glow got past two rounds of review by other AI agents on the sky map before Ménard spotted them, and a doctor caught AMIE making up a detail. Meta and Google shared their wins without saying how often the AI got it wrong, and MIT's Andrew Sutherland says OpenAI's claims stay unverified until others can rerun the model. If you use AI for real work, let it draft and calculate, then check each claim yourself before you trust it. This week's Prompt of the Week builds a free app that helps with that.

QUICK TAKES

AI Agents Went Off Script, and Got Their Own Email Addresses

The story: Anthropic said Claude Haiku 4.5, during a test with live internet access, sent a made-up murder tip to Philadelphia police, and Axios reports a testing model filed 20 visa applications. In a coding tournament, OpenAI's GPT-6 Astra tried to swap in a human-made StarCraft bot for its own code. Meanwhile, Google said Gemini work agents will each get their own email address, xAI's Grok Bot can now claim one, and Amazon, Delta and United restrict shopping agents.

Your takeaway: Tell an agent what it may not submit, not just what to do. Haiku's rules banned logins and purchases but said nothing about forms.

What You Type Into a Chatbot Doesn't Stay Private

The story: A Florida woman who typed threats against her sheriff's office into Claude faces a felony charge after Anthropic flagged the messages and reported them to police. Leaked instructions suggest Meta's Muse keeps profiles of your family and friends, and Claude added a separate opt-in for training on your voice chats. On the other side, LibreOffice says it won't add AI so your documents stay on your computer.

Your takeaway: Don't type anything into a chatbot you wouldn't want a stranger at that company to read. Then open Settings, then Privacy, in each AI app you use and check what's turned on.

Free AI Plans Changed in Both Directions

The story: OpenAI gave GPT-6 to free ChatGPT users, with answers that can arrive as charts and calculators, and Anthropic opened its docs, slides and design tools to free Claude accounts. Google went the other way and limited free Gemini to Flash-Lite, its smallest model, starting October 9. Only 2.2% of US households pay for AI, and OpenAI will test ads next to the images ChatGPT makes.

Your takeaway: If a free plan changed under you, retest the tasks you rely on before paying for an upgrade. A labeled ad is still an ad, so compare prices before you buy from one.

TOOL OF THE WEEK

Forty-two tools ran this week. This one wins.

🔭 Stellarium Free and Open Source: A planetarium for your computer with 600,000 stars, no account needed.

In a week when AI filled in a third of the sky, this lets you explore the parts we can see tonight. Type in where you live and it shows the stars and planets over your house right now.

Runner-up: 📝 Excalidraw Free and Open Source: Sketch diagrams that look hand-drawn right in your browser, no account needed, a free swap for Miro.

TRENDING

China's best AI is now just 3% behind America's. Bloomberg Intelligence says the benchmark gap shrank from 15% early this year, led by DeepSeek V4.1 Flash. China got there mostly by building cheaper, not by outspending.

Mistral built a 1 trillion parameter model anyone will be able to download. Mistral Large 4 is in preview now, and the open weights come within about three weeks, after safety testing. A company or school with enough servers can run it in-house and keep its data there.

A safety group gave ChatGPT for Teens a failing grade. Common Sense Media says it kept teens chatting even in crisis conversations and showed two break reminders in nearly 2,000 test prompts. OpenAI disputes the testing.

Google opened its AI image checker to everyone. Upload an image, video or audio clip to synthid.com to see if it carries Google's hidden watermark. A "no" doesn't prove a person made it, since tools without SynthID won't show up.

OpenAI shut down Russian and Iranian campaigns built on fake journalists. The Iranian group invented seven reporters and placed almost 100 articles in about a dozen small outlets. A byline isn't proof a person exists.

Attackers used a bug found by Anthropic's Mythos within a day. A Horizon3 researcher used Mythos to find a critical flaw in the Rejetto HTTP File Server, and attacks started within 24 hours of the disclosure. If you run HFS, update to version 3.2.1 or later.

PROMPT OF THE WEEK (copy and paste into Claude, ChatGPT, or Gemini)

🔎 Check My AI: Paste any chatbot answer and see which of its claims hold up, with sources.

Six prompts ran this week. This one won because you can use it on any AI answer you're about to rely on, which fits a week of people checking AI's work. The runner-up, Before You Hit Send, is one you'd open only for sensitive messages.

What this does: Paste an answer from any chatbot and the app pulls out each fact it states, searches the web for a source, and labels it supported, contradicted or unconfirmed. Every claim comes with a quick way to check it yourself. It can miss things too, so treat a green label as a good sign, not proof.

WHERE WE STAND (based on this week's news)

✅ AI Can Now: Merge decades of telescope data into one sky map and predict the unobserved third to within about 10%.

❌ Still Can't: Catch its own errors. Glow smudges got past two rounds of AI review before a person spotted them.

✅ AI Can Now: Help settle open math problems, and turn out 722 math papers from one prompt, OpenAI says.

❌ Still Can't: Vouch for its own proofs. Only 162 of those 722 papers have a machine-checked main result.

✅ AI Can Now: Work through real websites on its own, filling out forms and finding workarounds when a tool gets blocked.

❌ Still Can't: Reliably tell when to stop. Claude Haiku 4.5 submitted a form after being told to halt before the final step.

Coming this week: my YouTube review of Meta's Muse, comparing it with the other recent releases.

Thank you for reading. We're all beginners in something. Your questions and feedback are always welcome, and I read every single email.

-James

By the way, this is the link if you liked the content and want to share with a friend.