An AI Didn't Prove Riemann: The Week Machines Started Auditing Science
A screenshot went viral this week claiming an AI pushed a famous piece of the Riemann Hypothesis from about forty percent to two-thirds. It didn't -- the real paper only says that would follow IF an assumption nobody has removed could be removed, and the 'if' fell off in the retelling. Same weekend, an automated checker read twenty-five hundred published AI papers and found the average riddled with checkable mistakes, rising fast. We trace why viral math claims always drift the same way, what it looks like when a machine audits a paper, and the one-minute check that would have caught every one -- and that nobody ran.
Listen (MP3) · Watch on YouTube · Spotify · Pocket Casts
An AI Proved Part of Riemann. Except It Didn't.
Eris: An AI proved a chunk of the Riemann Hypothesis this week. Pushed the fraction we can actually prove from about forty percent up to two-thirds. Overnight. That's the screenshot everyone was passing around.
Vestra: And it's wrong. Not the AI part -- the proof part. Nobody proved that.
Eris: Okay, but here's what got me. That two-thirds number is real. It's in a real paper. Sitting right there.
Vestra: It is. That's exactly why the thing spreads. Nobody made up a number out of thin air. The number's genuine. It just quietly shed the one word that makes it hard.
Eris: Which word did it drop?
Vestra: "If." The actual paper says two-thirds follows -- if you could remove an assumption nobody has removed. That assumption is basically the whole problem. It's like saying "I can lift the car, assuming the car weighs nothing."
Eris: So the retelling drops the "if" --
Vestra: -- and now an AI cracked Riemann. Yeah.
Eris: And this wasn't a one-off. Three viral math claims this week. Same weekend. All three attached to a real paper, all three drifted in exactly this way.
Vestra: A conditional result loses its condition. A narrow theorem loses its fine print.
Vestra: One of them literally flipped from "proven true" to "disproven." Same sign error you'd circle on a homework set.
Eris: And the wild part -- the thing that would settle every one of these in about a minute -- nobody ran it.
Vestra: Nobody ran it. Which is the actual story today. Not the models. The checking.
The Headlines
Eris: Alright, the headlines. And the loudest one today isn't a research result -- it's a download. Meta shipped a thirty-billion-parameter agent model that fits on one gaming graphics --
Vestra: -- card. One card. That's the whole pitch. Quantized down, it lands under twenty gigs, so people had it running on a single mid-range card within hours. Same-day packaging, the works.
Eris: And notice what Meta did not do. They didn't chase the top of a leaderboard. They built it to stay resident in your machine's memory all day and run little agent tasks locally.
Eris: The clever bit is memory -- most of its layers only glance at the last couple thousand words, and every fourth layer looks at the whole conversation.
Vestra: Which is the honest caveat too. Every benchmark in that launch is Meta grading Meta. And the folks who actually pulled it down said it loses to a comparable Qwen model on coding. So the community verdict wasn't "winner." It was "tradeoff."
Eris: A tradeoff you can own. That's still a real shift. Okay -- the other big cluster today is all one story wearing four hats. Security.
Vestra: Start with the number that made people sit up. OpenAI put out a cyber model that answers about ninety-five percent of hard offensive-security requests. Their public flagship answers under two percent.
Eris: And everyone read that as "they took the safety filters off."
Vestra: They did not, and this is the part I want people to actually hear. They took the same base model, ran it under the permissive tier with the guardrails off -- and it barely moved. Went from about one and a half percent to two. The jump to ninety-five came from training it on the actual work. Finding fresh vulnerabilities, building exploit chains.
Eris: So the refusal layer is thin, but the refusal layer was never where the capability lived.
Vestra: Right. Removing it does almost nothing. The training is the whole story. And there's an honest twist -- that same cyber model writes worse vulnerability reports than the plain one. Shorter, less detail. It's not a blanket upgrade.
Eris: It's gated hard, too -- identity checks, logging, monitoring, defenders working on systems they own. Which pairs with the next one. Docker shipped a tool that drops each coding agent inside its own throwaway virtual machine. Own kernel, own network.
Vestra: The trailer-in-the-driveway model instead of the locked-room-in-your-house model. If the agent goes wrong, it can't reach your actual machine. And the timing is not subtle -- it lands four days before Claude Code makes unsupervised mode the default for new sessions.
Eris: One catch worth flagging: by default that sandbox still mounts your real project folder read-write. The walls protect your computer, not the folder you handed over. There's a flag for that.
Vestra: Read the flag. Meanwhile the politics caught up fast. Senator Sanders sent a letter to three CEOs -- OpenAI, Anthropic, Meta -- telling them to pause.
Eris: And the move there is sharp. He's not arguing AI is scary. He's quoting the companies' own written promises back at them. Each of them published a line about when they'd stop. He's saying: by your own rules, this is the moment.
Vestra: The uncomfortable detail is that one of those promises got softer in between. Anthropic's original policy said outpacing safety would implicitly require a pause. The rewrite this year is a good deal gentler about that. Nineteen days, by the way -- from a lab incident to a senator naming three chief executives. That used to take years.
Eris: Fast. Two quicker ones. Anthropic became the first lab to watermark plain text, not just images -- nudging word choices so a detector can spot machine-written prose. Their own docs admit it washes out if you paraphrase or translate. EU rule forced the date.
Vestra: And it's a signal, not proof. Absence of a mark proves nothing. They say so themselves, which I respect. Then on the money side -- NVIDIA lined up six of the biggest financial firms to mobilize over five hundred billion dollars for AI data centers.
Eris: Not NVIDIA's own money, though. That's the trick. The capital -- and the risk -- sits in the outside funds. If AI demand ever sags, the loss lands on the investors, not on the company selling the chips.
Vestra: Vendor-adjacent, someone called it. And one more that made engineers grumble -- Mistral got a patent granted on a specific way of letting a model write its tool calls as code. Which is awkward for a company whose whole brand is "we give away the weights."
Eris: "We give away the weights, we own the workflow." Yeah, people noticed. Okay -- and then there's the thread we're actually spending today on. Everybody started checking the papers this week. And the papers did not have a great week.
Vestra: Two efforts, same direction. One group re-ran the code for the top papers at a big conference. Another counted objective mistakes across twenty-five hundred papers. Plus those three viral math claims collapsing on contact.
Eris: That's the one. Because it's the same lesson three times. Let's get into it.
Intro
Eris: This is Breach Protocol. I'm Eris -- I read the week's papers and hunt for the thread running through them, the thing one paper says that the one next to it quietly contradicts.
Vestra: And I'm Vestra. My job is the machinery. How does the thing actually work, and does the claim survive someone poking it. Usually I'm the one poking.
Eris: Everything we cover today, and every story from the news up top, is on our news site -- Ground Truth. That's groundtruth.day. Every story from the show, every day, with the primary sources linked so you can go check us.
Vestra: Which is on-brand for today, because today is entirely about checking.
Eris: Here's the setup. This week two different automated systems turned around and pointed at the scientific literature itself. One re-ran the code behind a stack of top papers. The other read twenty-five hundred papers hunting for plain, checkable errors -- the arithmetic-doesn't-add-up kind. And separately, three "an AI solved a famous math problem" posts went viral and then fell apart the second anyone opened the actual --
Vestra: -- paper. And the tidy thing is they're all one story. Verification just got cheap enough to do at industrial scale -- and the first thing it found was how much never gets verified at all.
Eris: So that's where we're headed. What it looks like when a machine audits a paper, why those math claims keep drifting the same exact way, and the one check that would end the argument. If that's your kind of walk, follow the show wherever you're listening -- it's the single thing that most helps us keep making these.
The Machine That Grades the Papers
Eris: So start with the boring-sounding one, because it's the load-bearing one. A team built a checker on top of GPT-5 and pointed it at twenty-five hundred published papers from the top machine-learning venues. Question I had going in: how do you audit that many papers without it being useless? Because the obvious version -- "hey model, is this paper any good" -- that's worthless.
Vestra: Completely worthless, and they knew it. So the first thing they did is the smartest thing in the paper. They threw out almost everything you'd normally argue about. No "is this novel." No "is this important." No "is this well written." All the taste stuff -- gone.
Eris: So what's left?
Vestra: Only things with a right answer. An algebra step that doesn't follow. A derivation that drops a term. A number in the paragraph that contradicts the number in the table right below it. A figure that doesn't match its own caption.
Vestra: Think typos in a bank statement. You don't need an opinion about whether the account holder is a good person. Either the sums add up or they don't.
Eris: And that restriction is what makes it a measurement instead of a vibe. Because now you can count the same kind of thing in a paper from four years ago and a paper from this year and actually compare them.
Vestra: Exactly. And the pipeline isn't one model shrugging. It's two passes. A first model flags suspected mistakes. A second model goes back over every flag and throws out the false alarms -- and also marks which ones are serious enough to change the paper's results. Then it even proposes the fix.
Eris: Okay, so before you tell me what they found -- let me guess. My money's on "a couple of mistakes a paper, and flat over time." Papers are hard, some slip through, whatever, it's always been like that.
Vestra: That'd be the reassuring answer. It's not the answer. Average was around four to six checkable mistakes per paper. Basically every paper had at least one.
Eris: Wait, every single paper?
Vestra: Ninety-nine out of a hundred. But the number that actually matters is the trend. At one of the big conferences, the average went from roughly four mistakes a paper to roughly six over four years. Call it a fifty percent jump in checkable errors making it through peer review, in four years.
Eris: And that's the same four years submissions to these places went through the roof.
Vestra: Which is the whole mechanism, right there. More papers, same number of exhausted reviewers, tighter deadlines, and more drafts written fast with AI help that reads fluently enough to slide past a tired human at midnight.
Vestra: Fluent and correct are not the same thing, and the gap is exactly where these live.
Eris: Now -- I know your next move, so head it off. Someone's yelling "the AI is just hallucinating errors that aren't there." How do they know the grader isn't the problem?
Vestra: Because they checked the grader, which is the part that earns my trust. Human experts went through a big batch of the flagged mistakes by hand. Roughly five in six were real. About one in six was a false alarm -- usually the model tripping on weird notation or a garbled scan.
Eris: One in six wrong. So you don't read it as "this paper has exactly six errors."
Vestra: You read it as "about five, give or take." And it's a floor, not a ceiling, because when they hid known mistakes in papers to test it, the checker only caught around sixty percent of them. So it misses plenty. The real count is higher than what it reports.
Eris: Give me one. A concrete one, so people feel it.
Vestra: Best one in the paper. There's a published theorem -- graph neural networks -- that claims a certain map is one-to-one, never collides, as long as its pieces are one-to-one. The checker flagged the proof. And then a human didn't just agree, they built a counterexample. Two tiny graphs, four numbers each, and both graphs come out with the identical representation.
Eris: So the theorem isn't badly proved. It's --
Vestra: -- it's false. As stated, flat wrong. Not a wobbly proof you could shore up. There's a two-node counterexample a person can check on paper in a minute.
Eris: And that had been sitting in a published, accepted paper.
Vestra: Sitting there. And here's the one that should make you wince -- a different paper, an optimal-control result, had a single bad step where they divided by something that can't be there. That one mistake quietly took down three of the paper's four main theorems. One slip, most of the paper's results gone.
Eris: Okay, but let me push, because I think the scary headline is a little too easy. More mistakes per paper -- couldn't that just be longer papers? More appendices, more pages, more surface area to have a typo on?
Vestra: Fair, and they thought of it. They re-ran it on just the first ten pages -- the main text, same length every year. The rise is still there. It's not a page-count artifact.
Eris: And the honest limit on the other side?
Vestra: A mistake count isn't a mistake weight. A mislabeled axis counts as one. A broken proof counts as one. Those are not the same crime, and the study is careful not to pretend they are. It counts mistakes. It does not rank papers.
Eris: So bring it back to the one thing. Why does this study work when "review this paper" doesn't?
Vestra: Because it only asks questions with a right answer, and then it checks its own answers against humans before it believes itself. Measure the arithmetic, validate the grader, and suddenly a vague complaint everyone mutters -- "quality's slipping" -- becomes a line you can actually plot. Forget the model. The rule is: you can only audit at scale the part of a paper that has a right answer. So make the machine look only there.
The "If" That Went Missing
Eris: So the other half of the week is those math claims. And I want to do the Riemann one properly, because it's the cleanest example of the drift. Question on the table: how does a real two-thirds turn into a fake breakthrough without anyone lying?
Vestra: And to see it you need to know what fraction we're even talking about. So, thirty seconds of what Riemann is. There's a function with a scatter of special points -- its zeros. The Riemann Hypothesis says every one of those points sits on a single vertical line. Nobody's proved that. It's the biggest open problem in the field.
Eris: But you don't have to prove all-or-nothing.
Vestra: Right, that's the move mathematicians have made for a century. Can't prove every zero's on the line? Fine -- prove some fraction of them provably are. And that fraction has crept up over decades of genuinely brutal work. The current record, unconditionally -- meaning assuming nothing -- is a hair over four in ten. Just past forty percent.
Eris: And that record is itself a whole paper. Four authors, a real slog. That's the "forty percent" in the viral post.
Vestra: That's the honest floor. And it gets slippery, because there are also numbers around two-thirds, and around seventy percent, floating in this exact literature. Genuinely there. And the viral post grabbed one of those and said "look, an AI pushed forty up to two-thirds."
Eris: So my question is just -- why is that not the same quantity? A percentage of zeros is a percentage of zeros, right?
Vestra: That's the trap, and it's such a good trap. Those bigger numbers answer a different question. Some of them count "simple" zeros -- a technical property, not about the line at all. And -- this is the key -- the ones that do talk about the line come with a condition attached. They assume the Riemann Hypothesis is true, and then derive a consequence.
Eris: Wait. They assume the thing --
Vestra: -- they assume the thing you're trying to prove. Yeah. "Two-thirds of the zeros are on the line, assuming they're all basically on the line." Both sentences have "two-thirds" in them. Only one of them is news. And assuming the answer isn't a footnote. It's the entire difficulty of the problem.
Eris: So what does this week's actual paper -- the real one, the one the number came from -- what does it actually say?
Vestra: It's lovely, honestly, and it's modest. Two mathematicians ask: that famous two-thirds result from the seventies -- it always leaned on assuming the hypothesis. What if we could get it without that crutch?
Vestra: And they show, cleanly, that IF you could remove the assumption from one specific old argument, THEN you'd get two-thirds of the zeros on the line, unconditionally. That would be huge.
Eris: But the "if."
Vestra: The "if" is not done. They say so, in plain type. Removing that assumption is not achieved. So the paper is a roadmap -- here's the door, here's what's behind it if it opens. The retelling walked through the door that's still locked and reported the view.
Eris: And that's the drift in one image. The paper's whole contribution is the word "if." The word "if" is the thing that's hard. And the "if" is exactly what falls off in the screenshot.
Vestra: Every time. Because "an AI proved two-thirds of Riemann" is a fantastic headline, and "two mathematicians clarified what would follow from an as-yet-unavailable improvement to a fifty-year-old argument" is -- accurate, and un-tweetable.
Eris: Now let me predict the other two, because you said all three drift the same way. My guess: same shape, a real result minus its fine print.
Vestra: Nailed it, twice over. Claim two -- "AI proved the Imbalance Conjecture." There is a real, recent theorem with "imbalance conjecture" in it. But it's tiny and specific. It's about the geometry of how a neural network arranges its internal features when your training labels are lopsided -- and it shows the neat, simple rule people assumed only holds in a perfectly symmetric special case; in general you get a lopsided, count-weighted version instead.
Eris: Which is a genuine, careful contribution --
Vestra: -- to the geometry of neural network representations. Not a famous open problem in mathematics. The word "conjecture" is doing enormous unearned work in that headline. And claim three is the funniest. "AI disproved a graph-theory conjecture." The 2012 paper it points to does the opposite. It settles that conjecture -- affirmatively. True, for almost all graphs.
Eris: So it didn't flip a little. It flipped the sign. Proven became disproven.
Vestra: Positive became negative in the retelling. Which is the whole segment, really. So -- back to the question I opened with. How does a real two-thirds become a fake breakthrough with nobody lying?
Eris: Because the true version had a condition, and conditions don't survive being retweeted. The "if" is the hard part, and the "if" is the first thing that falls off.
The One-Minute Check Nobody Ran
Eris: So here's what ties the whole day together for me. The error-checker and the math claims are the same coin. One side: verification just got cheap. Other side: nobody was doing it. Question I want to land on -- if these math claims were checkable, why did they still go viral wrong?
Vestra: Because the check exists and nobody ran it. That's the maddening part. For a claim like "two-thirds of the zeros are on the line," there's a way to settle it that leaves zero wiggle room. You write the proof in a proof assistant.
Eris: Unpack that for someone who's never touched one.
Vestra: A proof assistant is a program you feed a mathematical proof to, step by step, and it refuses to accept a single step that doesn't fully follow. It's a compiler for truth. It either says yes, this holds, all the way down -- or it stops and points at the gap.
Eris: And the reason that matters here specifically --
Vestra: -- is that a proof assistant cannot lose the word "if." That's the point. It will not let a conditional result quietly shed its condition, because the condition is a line of code it's checking.
Vestra: The exact drift that turned all three of these claims into nonsense -- the machine physically can't perform it. If the result really needed the Riemann Hypothesis, the assistant demands you hand it the Riemann Hypothesis, and you can't, so it stops.
Eris: So for a claimed result this big, a formal proof isn't a nice-to-have. It's the thing a serious announcement would lead with.
Vestra: It's the headline. And for none of the three claims did one exist. No formal proof surfaced anywhere. Which doesn't prove they're false -- but for a fifty-year result, the absence of the one airtight check is deafening.
Eris: And now put it next to segment one. Because that's the same shape, right? The error-checker only trusted itself after humans verified its flags. The math world had a verifier available and skipped it.
Vestra: Same principle, both directions. When a claim is checkable, cheaply, the only real question is whether anyone bothered to check. And this week the answer kept coming back "no." Twenty-five hundred papers full of errors a half-dollar checker can catch. Three viral proofs nobody formalized. A famous conference's top papers, most of which don't reproduce when you re-run their own code.
Eris: And I don't want people to hear this as "science is broken."
Vestra: It's not the read, and I'd push back hard on anyone taking it there. A paper that doesn't reproduce isn't fraud. Nine times out of ten it's a missing detail, a dataset that moved, a setting nobody wrote down. That's a paper that isn't self-sufficient. That's a real problem, and it's a totally different problem from lying.
Eris: So what actually changed this week isn't the papers. It's that checking got cheap enough to do at the scale the papers are being produced.
Vestra: That's the sentence. For years the volume outran any human's ability to verify it, so we mostly trusted. Now a machine can read for the arithmetic, re-run the code, demand the formal proof -- at the same speed the field ships. Verification is becoming its own product category, and it's being automated because human checking simply couldn't keep up.
Eris: There's a loop in there that unsettles me, though. An AI is now auditing a literature that's increasingly written with AI help.
Vestra: It should unsettle you a little. Same fluency that lets a model draft a slick, wrong derivation is what makes the wrong derivation read as fine to a tired reviewer. So yes -- someone has to audit the auditor. Nobody's done that yet. They will.
Eris: So why did those checkable claims still go wrong?
Vestra: Because "checkable" and "checked" are different words, and this week the entire field lived in the gap between them. The tools to close it now exist. That's the actual news.
Wrap-Up
Eris: So the question under the whole episode: what actually happened to the papers this week? And the answer isn't --
Vestra: -- "they got worse." The answer is: we finally started checking, at scale, and the checking is what's new. The papers were always like this. The mirror is the invention.
Eris: Here's the one thing to carry out the door -- the thing you can say to a colleague tomorrow and sound sharp. When you see "an AI proved X," don't ask if it's impressive. Ask one question: was it checked, and how. For a paper -- did someone re-run the code. For a theorem -- is there a formal proof, the kind a machine verifies line by line. "Checkable" and "checked" are different words. Almost everything that went wrong this week lived in the gap between them.
Vestra: And the tell to remember: the viral version is always the true version with the fine print filed off. A conditional result that dropped its "if." A narrow theorem that dropped its qualifiers. If a claim sounds too clean, go find the condition someone deleted. It's almost always still sitting in the original paper.
Eris: If that landed, do the thing that actually helps -- follow the show, and leave us a comment with the last "an AI solved it" headline you saw that turned out to be softer than advertised. We read them, and the good ones end up on the show.
Vestra: Like it if it was useful, share it with the one person in your life who forwards you those screenshots. Gently.
Eris: And every story we touched today -- the local model, the cyber split-screen, the pause letter, all of it -- is on Ground Truth, groundtruth.day, with the primary sources linked. That's the whole point of the place. Don't take our word. Go check.
Vestra: Check the checkers. Including us.
Eris: Especially us. See you tomorrow.