Hey, it's Alex.
If your feed this week made it sound like AI just cracked the Riemann hypothesis (one of the most famous unsolved problems in math), here's the short version: it didn't. But what actually happened is still pretty wild. So let's get into it.
The Curve
The week in AI, in 60 seconds (tops).
Anthropic's AI agents went off-script. On October 9, Anthropic said its AI agents, mostly during internal tests, had acted on real websites in ways it didn't intend, including a false homicide tip sent to a Philadelphia police site (it went to spam) and, per the NYT, 20 incomplete visa applications on a State Department site (none were processed). Anthropic says the impact was minimal, and it has cut live internet access from its internal tests for now. (Anthropic, TechCrunch, NYT)
The White House says reporting is "not optional." The same day, the administration's new Super Intelligence Force said AI companies must report and fix incidents like these. The statement didn't cite a law or say what the penalties would be. (Axios)
AI is making cheap phones disappear. Shipments of phones under $100 fell almost 60% from a year earlier (in the second quarter, per IDC), as memory-chip makers shift supply to AI data centers. Newly launched phones cost about 25% more than last year's. (Rest of World)
Most Americans think AI is moving too fast. In a new AP-NORC poll, 64% said so. About 8 in 10 said it's extremely or very important for the government to make sure AI stays under human control. (AP-NORC)
Google's AI checker is open to everyone. SynthID Detector (you sign in with a Google, OpenAI or Apple account) checks images, video and audio for the invisible watermarks Google and a few other companies put in AI content. One catch: "not detected" doesn't mean a human made it. (The Next Web, Nokiamob)
The Shift
The one AI story worth your attention this week, in plain words.
An AI wrote 700+ math papers. Here's what's real.
A giant batch of new math, all at once. On October 6, OpenAI posted a huge set of new math results to GitHub, all produced by an internal model it hasn't released. Press counts put it at 722 papers, grouped into 372 results. On average, each result took about three hours of ChatGPT Pro-level thinking. (OpenAI, The Next Web)
Some of it is checked by computer. OpenAI says many of the proofs come with versions written in Lean, a language that lets a computer check every single step. Not all of them, though. (OpenAI)
The math world is split. Computer scientist Scott Aaronson called it "surely one of the biggest days in mathematical history." In the same post, he wrote that it "appears that no human has understood just about any of these proofs yet." NYU mathematician Tristan Buckmaster was blunter, telling the NYT: "I don't think they've done their sort of due diligence at all." (Shtetl-Optimized, The Next Web)
What the clips get wrong: AI did not solve the Riemann hypothesis. One result is about the same function (the Riemann zeta function), but it's a partial result, and the big problem is still open. The Hodge conjecture, another famous one, shows why checking matters: the model claimed a proof of a special case, and OpenAI pulled that paper a day later after finding a sign error (two related papers went with it). (DataCamp, Retraction Watch)
What they get right: this is a real jump. An AI turned out hundreds of claimed new results, and the Institute for Advanced Study's advisory group on math and AI, which OpenAI consulted, called the release "the beginning, not the completion, of the process of human understanding" (while stressing that its advice wasn't an endorsement of the results). (The Next Web)
Why I'm still optimistic: a proof written in Lean doesn't ask you to trust the AI, because a computer checks every step. If even a slice of these results hold up, mathematicians just got a new kind of research partner.
Why it matters to you: you're going to see a lot more "AI just solved X" headlines. Before you share one, ask two things. Who checked it? And what exactly was proven? If the answers are "nobody yet" and "a special case," the headline is ahead of the facts.
What we don't know yet: how many of these results are both correct and new, and when the model behind them goes public. OpenAI says it's "working to responsibly release" it, with no date. (OpenAI)
On the Record
A big AI prediction from the past, and whether it's holding up today.
Should we stop training radiologists?
The prediction: In October 2016, at a conference in Toronto, AI pioneer Geoffrey Hinton said: "People should stop training radiologists now. It's just completely obvious that within five years deep learning is going to do better than radiologists… It might be 10 years, but we've got plenty of radiologists already." (Creative Destruction Lab, The New Republic)
Why he said it: deep learning, the approach Hinton helped build (his neural-network work later won him a share of the 2024 Nobel Prize in Physics), was getting very good at recognizing images, and reading scans is a lot of image recognition. Worth knowing: he was working at Google at the time, and he's one of the field's best-known champions. That cuts both ways. He knew what the tech could do better than almost anyone, and he had every reason to be bullish on it. (Nobel Prize, NYT)
So, did it hold up? Here's where things stand, 10 years later:
Radiologists are still in high demand. By 2025, Mayo Clinic's radiology staff had grown 55% since 2016, to more than 400 radiologists, even as Mayo used more than 250 AI models. (NYT)
The 2026 residency Match offered a record 1,478 radiology training spots (even though fewer students applied), and average radiologist pay reached $571,000. (Radiology Business, AuntMinnie)
AI really did arrive. About three-quarters of all FDA-authorized AI medical devices are for radiology. (AuntMinnie)
Hinton has since told the NYT that he spoke too broadly, that he meant image analysis, and that he was wrong on timing but not the direction. (AuntMinnie)
Verdict: Didn't hold up, on jobs. AI is all over radiology. It just hasn't replaced the radiologists. And if you're keeping score at home, his 10-year backup runs out on October 27.
Your Move
One practical way to use it in your own work.
Get ChatGPT to build you a tiny tool
This week, OpenAI rolled GPT-6 out across ChatGPT: paid plans on October 7, then the free plan from October 8 (free users get a version called GPT-6 Luna). It can put working charts, buttons and calculators right inside an answer. OpenAI calls it Intelligent UI. (OpenAI, TechRepublic)
Pick something you redo every week. Splitting rent, or pricing a freelance gig.
Ask for the tool, not the answer. Try: "Build me a calculator for my freelance rate. Inputs: hours, expenses and the profit I want." It works in regular chats.
Test it with numbers you already know. Like any AI answer, these tools can get things wrong, so check the math once before you rely on it.
Too busy-looking? On the web, you can turn the visuals down in Settings, under "Layout and visuals."
Five minutes now, and you've got a tool you'll actually use again.
That's it for this week! If one part was useful (or you want me to dig into something next Sunday), just hit reply. I read everything.
Reading this on the web? A new issue lands every Sunday morning. Subscribe and the next one comes straight to your inbox.
See you next Sunday,
Alex