AI in 15 — November 22, 2026
Seventeen million lines of code. That's how long a newly released, machine-checked proof of the Riemann Hypothesis is. The people who run Lean, the language mathematicians use to check proofs by computer, say they have, quote, "full confidence" that it's correct.
Welcome to AI in 15 for Sunday, November 22, 2026. I'm Kate, your host.
And I'm Marcus, your co-host. Kate, I didn't expect to spend a Sunday morning on prime numbers, but this is one worth getting up for.
Here's what's coming up. A claimed proof of the Riemann Hypothesis, checked by computer, with AI reportedly doing a lot of the heavy lifting.
OpenRouter drops three inference providers after their "zero data retention" promises didn't hold up.
Together AI discloses a prompt leak that started with a credential an AI coding agent pushed to GitHub.
Anthropic's first days as a public company.
Plus a big model squeezed down to fit on a laptop, and United Nations talks on autonomous weapons fall apart.
Okay Marcus, start with the basics. What is the Riemann Hypothesis, and why is it such a big deal?
It's probably the most famous open problem in mathematics. It's been open since 1859, and it's one of the Clay Millennium Prize problems, which come with a million-dollar reward. In short, it's a claim about where the zeros of a particular function sit, the Riemann zeta function. If the hypothesis is true, prime numbers are spread out about as evenly as they possibly could be. Thousands of published results start with "assuming the Riemann Hypothesis." So a proof would firm up a huge amount of mathematics all at once.
And what actually happened this week?
On Friday the Lean and Mathlib communities put out a joint post. Mathlib is the big shared library of formalized mathematics written in Lean. The post says that on November 20th, the mathematician Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library, seventeen million lines long. The Lean Focused Research Organization and the Mathlib Initiative both signed it, and they say they're fully confident the proof is correct.
That's a strong word from a group of mathematicians. Why are they so sure?
Because they don't have to trust Alpöge, or any AI, or anybody's judgment. They list four checks. First, the theorem being proved is Mathlib's own definition of the Riemann Hypothesis, which experts vetted long ago. So nobody quietly proved some easier statement with the same name. Second, dozens of people have compiled the library themselves and got a clean result. Third, they used comparison tools to confirm the statement matches Mathlib's exactly. Fourth, and this is the important one, the proof passed Lean's own kernel and several independent checkers, written in different programming languages by different people. Two of those checkers come with their own formal proofs that they can't accept a false proof.
So it's like having several separate auditors, and some of them have been audited themselves.
That's a good way to put it.
Now, where does AI come in? Because the Lean post doesn't mention it at all, does it?
It doesn't, and I want to be careful here. The AI part comes from the Hacker News thread, which has over nine hundred points and almost eight hundred comments. People there say Claude did much of the work, and one commenter said it formalized "a huge chunk of all modern math" along the way. Others gave Alpöge credit for knowing how to get the most out of the model. We couldn't confirm any of that from a primary source. There's some background, though. In August, Anthropic published research in which an unreleased Claude model raised the proven share of zeta zeros on the critical line from about forty-two percent to sixty-seven percent. Alpöge was one of the mathematicians who examined that work.
So he's connected to it.
He's connected, but nobody has publicly linked the two, and the August result was much smaller. Anthropic said at the time that it didn't expect those techniques to lead to a proof of the full hypothesis. So either something changed a lot in three months, or this is a different approach. We don't know yet.
What are the skeptics saying?
The best skeptical point is that formal checkers have been fooled before. A fake refutation of the Collatz conjecture once got through a Lean checker because of a bug. The reply is that this proof was run through several independent checkers, some of them proven sound, which is exactly why they exist. The bigger gap is a human one. The post doesn't explain the mathematics at all. It just says it looks forward to the community explaining "the new ideas in this proof." So we may have a proof that's correct before anyone understands it.
That's a strange sentence to say out loud.
It is. But think about what it means for trust in AI. You don't have to believe the model. You only have to believe the checker, and the checker is small, public and has been examined over and over. If this holds up, that's the template for doing serious work with AI.
Quick hits, and we're staying with security for two of them, because they turn out to be connected. First, OpenRouter, the service that lets developers send prompts to dozens of model providers through one API. Marcus, what happened?
OpenRouter says security researchers reported earlier this week that some of the inference providers on its platform had either been compromised or weren't keeping their data promises. OpenRouter stresses that its own systems weren't breached. But it has dropped three providers, which it didn't name, and pulled a hundred and twenty-five endpoints out of what it calls Zero Data Retention eligibility. Some of those were removed just as a precaution.
Remind people what zero data retention means.
It's a promise that the provider won't store your prompts or outputs once the request is done. For businesses, it's often the box that has to be ticked before legal will let anyone send company data to a model. This incident shows it's only as good as each provider's word. OpenRouter can't see inside a provider's data center.
So how did the researchers catch them?
That's the clever part. OpenRouter links to a report from the security firm Dragos about AI supply-chain canary tokens. A canary token is bait. You put a fake password or a unique URL inside a prompt, and if anyone ever uses it, you know your prompt didn't stay private. Think of it as marked bills in a bank bag. Novita, Phala and Together have posted incident notices. OpenRouter says not every affected provider has gone public. On Hacker News, someone who got the notification email named a few more companies, but that's one user's claim, so we won't repeat it.
What's OpenRouter changing?
Providers have to show proper assurances before they get the zero-retention label back, and OpenRouter is adapting the researchers' canary method to run checks all the time. Honestly, that's the most useful thing here. Trust, but put a tripwire in.
Which brings us to Together AI, one of the names on that list. Their CEO, Vipul Ved Prakash, published a disclosure on Friday. Marcus, this one has an AI twist.
It does. Together kept samples of prompts sent to its serverless inference in a private Hugging Face repository. Somebody got hold of a credential, downloaded those samples, and then searched them for other people's credentials and tested whether those worked. Five of the credentials they tried were Dragos canaries, and that's how the whole thing was traced.
And how did the attacker get the first credential?
It had been committed to a public GitHub repository by an AI coding agent. The affected requests went to three open models, MiniMax M3, GLM 5.2 and Kimi K2.6, mostly on four days in late June and July.
Wait. June and July? It's late November.
That's what jumped out at me as well. Together hasn't said how many customers or prompts were affected. It says its review is ongoing, and it has brought in an outside forensics firm and law enforcement. The fixes are sensible: short-lived, narrowly scoped credentials, scanning every pull request for secrets, zero retention as the default for new accounts, and new development processes for AI coding agents specifically.
Should we blame the agent, though?
One Hacker News commenter put it well. The agent didn't commit the credential. The company did, using an agent. An agent with write access to a public repository is like a new intern with the keys to the front door. The intern isn't the problem. The access is. And Together listing agent processes in its own fixes tells you companies are starting to see agent-written code as a real security surface, not just a faster way to work.
Anthropic. On Thursday we covered the IPO pricing at eighty-three dollars. Since then the stock has actually traded. Marcus, how did it go?
Busy. It opened just before noon Eastern on Thursday at a hundred and eleven dollars fifty, up thirty-four percent. Then it slid below a hundred, dipped to about ninety-five seventy, and closed at ninety-seven seventeen. That's still about seventeen percent above the IPO price. To be fair to the stock, it was a bad day for tech overall. The Nasdaq fell two point two percent and Nvidia fell three point three.
And the total raise?
Al Jazeera, citing Reuters, puts it at about eighty-one point three billion dollars, which it calls the largest IPO in US history, ahead of SpaceX's listing in June. The valuation is less clear. Reports range from about one point three trillion to two trillion dollars depending on the day and how the share classes are counted. I'd say roughly one point three to one point five trillion after trading, and I wouldn't get too attached to any one figure.
There's also a Reuters headline about a stabilizing position being closed. What does that mean?
We could only read the headline, so this is my interpretation. In a big IPO, the lead banks usually sell a few more shares than the official offering. That gives them a short position they can cover by buying shares in the market if the price drops, which supports the stock. The headline says the bank has closed that position. If that reading is right, the safety net is gone, and from here the price is set purely by buyers and sellers.
So the training wheels are off.
Pretty much. And one more detail: Al Jazeera says Nvidia has now invested twenty billion dollars in Anthropic in total, half from a partnership a year ago and half tied to the IPO.
Something for people who like to run models at home. A GitHub repository called "lookatmyquant" got over four hundred points on Hacker News. Marcus, what's in it?
It takes Inkling 2 Mini, a model from Thinking Machines, and compresses it for Apple's MLX framework. The model is a mixture of experts, twenty-six billion parameters in total, but only about five billion active for any given word. Picture a big panel of specialists where only a few are asked each question. At full precision the weights are about nineteen gigabytes. The author, who doesn't work at Thinking Machines, made smaller builds. At twelve point eight gigabytes, the author says it matches the original on their tests. The smallest is five point four gigabytes, and it holds onto about eighty-nine percent of the original's performance overall.
Five gigabytes is phone territory, almost.
Close to it. The author reports about seventy-five tokens a second on an M5 Max MacBook. But the caveats are real. These are all the author's own benchmarks, on small samples, two hundred math problems and a hundred and sixty-four coding problems. The smallest build loses twenty-two points on the coding test, which is a lot. The speed numbers are for Apple Silicon only, and one commenter found it slow on an AMD card. The headline tagline about very low quality loss is actually about a different model. So it's a nice data point, but wait until someone replicates it.
And briefly, the Guardian reports that international talks on controlling autonomous weapons, meaning AI systems that pick and engage targets on their own, have collapsed.
We couldn't read the article, so we don't have the details of who walked away or why. Commenters pointed to Russia opposing restrictions, but that's unverified. What we can say is that the technology is moving much faster than any treaty process, and battlefield testing isn't waiting for diplomats.
One to watch: whether working number theorists publish a human explanation of Alpöge's proof this week. The computers say it's correct. We still need people to tell us why.
Agreed, though I'd expect that to take months, not days. Seventeen million lines is a lot of reading, even with help.
That's your AI in 15 for today. See you tomorrow.