Hope Sterling: Hey you, ready for this one? I'm handing you a number and I need a reaction. Seventeen million lines of code. That's the proof.
Juniper Vale: Oh, wow. Seventeen million lines, for what, exactly?
Hope Sterling: The Riemann hypothesis! Bernhard Riemann posed it in 1859, so it's been sitting there unsolved for 167 years, million-dollar prize and everything. And on November 20th or 21st, Levent Alpöge, who says he did it with Claude, just posts a four-page note on his personal website. Like, it's an engagement announced in a Notes app screenshot.
Juniper Vale: I mean, four pages on top of 14 archive files of Lean code. That's the mismatch that gets me.
Hope Sterling: I did the math. One line a second, nonstop, that's like 197 days of reading. Literally over half a year!
Juniper Vale: And nobody reads it that way. So here's the click. Lean is a spell-checker. It confirms every step follows from the one before it.
Hope Sterling: Spell-checker, so, like, it catches typos but not whether the essay's any good?
Juniper Vale: Right, right, and that's the part that matters. It can't tell you why the argument works. That takes an editor, a person who understands it. For now we have the checked spelling and four pages sitting on 17 million lines.
Hope Sterling: So if nobody can read it, who even says it works? Because I saw headlines on Saturday saying it checks out. Did somebody actually check?
Juniper Vale: Two organizations did, yes. By Saturday afternoon, November 21, two independent groups announced they'd separately verified the Lean code. The Guardian and Scientific American quoted experts saying it checks out, that it's almost certainly true.
Hope Sterling: Stop. Then it's proven! Pay the man his million.
Juniper Vale: Not so fast. The Clay Mathematics Institute named it a Millennium Prize Problem in 2000, with that million dollars attached. Hours after the news, their site still said unsolved. And they require a two-year publication and peer-review period.
Hope Sterling: Two years?! Okay wait, so like, the code compiles, but the prize isn't won. Those are literally two different sentences.
Juniper Vale: Different sentences, yeah. And one more gap. Lean checks the theorem as it's written in Lean. So a person still has to confirm that the formal statement actually means the Riemann hypothesis. That piece is small enough for a human to read, and it matters a lot.
Hope Sterling: And Alpöge, he worked on this for how long? With Claude, I mean, like, is this a years thing?
Juniper Vale: About six weeks. He works at Anthropic, which built Claude. Now, reports on November 22 conflict. Some say verified, some say unconfirmed. I'll say it once: the status is unsettled. At 100 lines a minute, 17 million lines takes about 118 days. As of today, no human has read the whole argument.
Hope Sterling: Hold on. So peer review can't mean, like, reading it?
Juniper Vale: Right — and that's the surprise. It has to mean trusting the tools, then rebuilding the ideas some other way.
Hope Sterling: But wait, okay, so there's a guy at Stanford, Brian Conrad, who like, literally collaborated with Wiles, and he called this whole thing "frankly completely outrageous." Releasing a proof the author doesn't fully understand. And my first thought was, oh no, he thinks it's wrong?
Juniper Vale: No, he's not saying that. His objection is about what a proof is for. It's supposed to show you why something is true, not just that it is.
Hope Sterling: Ohh. So it's about the practice. Like, the job of math.
Juniper Vale: Right. Epistemology, if you want the fancy word. And Alex Kontorovich at Rutgers made a quieter version of it: results like this would normally show up in a formal paper, not as Lean code.
Hope Sterling: Okay, but here's my plot twist, and I cannot with this. Same day, Ravi Vakil, he's the AMS president, posts a totally different proof. Alessandro Masullo, 940 pages, all human. And Conrad had been reviewing that one for, like, months.
Juniper Vale: And it's the Generalized Riemann Hypothesis, a bigger statement. Then about eight hours later Vakil updated his post. Multiple problems flagged.
Hope Sterling: Stop. Flagged by who?
Juniper Vale: Mathematicians, and some of the problems were found using AI models. I'll be honest, the sourcing on the details is thin, so I wouldn't go further than that. But the shape is the thing: AI is the author of one proof and the proofreader of the other.
Hope Sterling: So nobody wins the like, humans versus robots bracket? Okay, and he doesn't fully understand it. Who even gets credit for that?
Juniper Vale: Hold that thought. Picture one number theorist opening those 14 files, and you'll see why it's the real question.
Hope Sterling: If nobody can read it, then somebody's gotta do the translating. Show me her. The number theorist.
Juniper Vale: Picture a number theorist at a university. Dozens of her published theorems start with the words "assuming the Riemann hypothesis." She downloads the 14 files, and her laptop fans spin up as the checker runs. Then a green checkmark.
Hope Sterling: Oh, the relief. Like, literally, all her papers just got a floor under them!
Juniper Vale: For about ten seconds. Then she asks the real question. Can I teach this to my grad student? Can I cite a step? And the answer is no, not yet. So she opens a blank document and starts writing, in plain English, what the proof might be saying.
Hope Sterling: Aww. Okay, but, like, is a checked answer nobody understands actually knowing something? Or is it just, I don't know, a really good receipt?
Juniper Vale: I'd say a receipt. It proves a purchase happened. It doesn't tell you what's in the bag. And that's why Levent Alpöge saying he doesn't fully understand it is the most remarkable fact here, and the most troubling. Credit between him, Claude, and Anthropic, who employs him and built Claude, is genuinely unclear. Traditionally authorship rests on who understood and constructed the argument.
Hope Sterling: Wait, so does the credit go to her? The one writing the blank document? Hmm. But then Reese Pathak, the Berkeley postdoc, says these systems solve in hours what he thought would take years, so like, is the slow part the bottleneck now?
Juniper Vale: Maybe the bottleneck moved, not vanished. Six weeks to make the thing. But a Stanford-style postdoc with a blank document, writing what 14 files might be saying? That part doesn't speed up the same way. Nobody's found a shortcut for it yet.
Hope Sterling: I keep picturing her. Tuesday morning, coffee going cold, and there's just the cursor. Blinking. And she has to type one sentence, like, in actual English, that starts with, I don't know, "The idea seems to be..." and that's literally the whole job. Aww. Wait, is that the discovery? The sentence?
Juniper Vale: Maybe. If the proof holds up, a person still has to explain it. Whether that's optional or the real discovery, I don't think anyone knows yet. I'd lean toward the real one.
Hope Sterling: Mm. Me too. Even Brian Conrad, he's not mad it exists, he just wants somebody to be able to say why.
Juniper Vale: That's the one thing I'd sign. The checkmark is real, the Clay Institute clock is still running, and the first plain sentence isn't written.
Hope Sterling: Okay, I can sit with that. Go pour her a fresh coffee, I guess. She's got a long week.