OpenAI AI Claims 700+ Math Solutions, Potentially Historic

OpenAI released a document claiming 700-plus mathematical "solutions," and Alberto Romero called it a civilizational inflection point, comparing it to a nuclear bomb for mathematicians. TechCrunch, the same week, reported the solutions aren't meeting the field's standards yet. Both are in the public record. Figuring out which signal to trust is the whole exercise.

Start with what's missing. OpenAI hasn't said what actually produced these results. That single fact decides everything downstream. Machine-checkable formal proofs in Lean or Coq could be verified in weeks and would represent a genuine capability leap. Natural-language arguments needing human expert sign-off are a release strategy, and could drag on quietly for a year before deflating. Until OpenAI publishes the method, you're reading a marketing document.

"At varying stages of verification" covers a lot of ground. It could mean 695 are garbage and 5 are incremental. The prior on one system generating hundreds of correct novel proofs simultaneously is low, and Andrew Wiles's proof of Fermat's Last Theorem, far more scrutinized than anything in this drop, had a serious gap surface months later. That's the base rate for deep mathematical results.

The asymmetry matters, though. If even three to five hold up at the claimed level, this is genuinely historic, and the 695 wrong ones stop mattering. That's why the story won't die quietly.

One underpriced concern: if this is fully real, it means humans have lost the ability to check the work at the speed it's produced. Magnus Carlsen can't beat or fully follow a chess engine anymore. The same dynamic applied to mathematical reasoning means the "human in the loop" safety story stops working, because you can't supervise what you can't verify in time.

The release format is itself evidence. Genuine Fields-level work goes to journals and proof-checkers because the author wants the stamp. A document drop wants the headline first and the verification later.

The call: By April 2027, the number of OpenAI's claimed solutions independently confirmed by established mathematicians as novel, correct, and of the claimed historic importance will be in the single digits. Hundreds confirmed is not what the evidence supports, and the field has never once verified work at that speed.

OpenAI dropped a document with 700-plus mathematical "solutions," and Alberto Romero is calling it a civilizational inflection point. Serious people are quoting him calling it a "nuclear bomb for mathematicians." TechCrunch, the same week, ran a quieter headline: the solutions aren't meeting the field's standards yet. Both things are in this cluster. Figuring out which one to believe is the whole exercise.

The thing being decided for anyone who builds with these models is simple. Is OpenAI's reasoning capability now good enough that you should re-price what you expect from it on hard, novel problems? Or is this a press release with a long tail of wrong answers and five good hits? That's easy to undo either way. Nobody has to sign anything this month. There's no deadline here except the verification clock, which runs in months and years, not days.

The Skeptic. "700 solutions at varying stages of verification" is a phrase built to survive being wrong. It could mean 695 are garbage and 5 are incremental. The "nuclear bomb for mathematicians" line, delivered with "no hesitation or qualification whatsoever," is the part that should worry you most. When people make unqualified civilizational claims days after a document drop, before anyone has checked the work, someone is being sloppy. The track record is not kind: AlphaProof's IMO run was real but bounded, GPT-4 face-planted on arithmetic it should have aced. Impressive math demos stall exactly where verification gets hard. The Navier-Stokes comparison is doing so much work it's holding the whole argument up by itself.

The Researcher. Verification is the entire story, and it hasn't started. OpenAI doesn't get to grade this. Romero doesn't. Twitter doesn't. The mathematical community does, and Fields-level results take years, not weekends. Andrew Wiles proved Fermat's Last Theorem and a serious gap turned up months later, from a proof far more scrutinized than anything in this drop. The prior on one system producing hundreds of correct novel proofs simultaneously is low. But the floor matters: if even three to five hold up at the claimed level, this is genuinely historic, and the 695 wrong ones won't matter at all. That asymmetry is why the story won't die quietly.

The Builder. Nobody is asking the question that decides everything: what actually produced these? Three completely different machines could sit behind this document. A reasoning model running long chains of thought over known conjectures. A formal proof assistant spitting out Lean or Coq that a computer can check in weeks. Or natural-language arguments that need human experts to adjudicate one by one. Those have wildly different reliability. Machine-checkable formal proofs would be extraordinary and settled fast. Natural-language "solutions" needing expert sign-off are a release strategy, not a verified result. Until OpenAI publishes the method, you're reading marketing, and you can't ship anything on top of a marketing document.

The Safety Lens. Strip the cultural hand-wringing and Romero's Carlsen-versus-Stockfish point is the real concern. Magnus Carlsen can't beat or even fully follow the chess engine anymore. The moment a model's math reasoning runs past the speed humans can check it, the "human in the loop" safety story stops working. You cannot supervise what you cannot verify in time. That's not science fiction, it's the concrete version of a known problem. And OpenAI shipping this as a document drop rather than through peer review puts announcement speed ahead of verification. For a company that has built its brand on careful oversight, that order is backwards.

The Compute Pragmatist. If a model genuinely generated 700 novel proofs of real depth, the cost of running that is extraordinary, and it tells you where the reasoning-cost curve actually sits. Long-horizon novel proof search is about the most expensive thing you can ask a model to do, far above coding or summarizing. So one of two things is true. Either OpenAI cracked dramatically cheaper reasoning by training the model against formal-verification feedback, which would reset inference pricing for everyone, or this is bulk generation: throw enormous compute at the problem, generate thousands of attempts, and cherry-pick the handful that land. The second is far more common and far less exciting. The method tells you which.

Where they split. The Researcher and the Skeptic agree the claim is unproven but disagree on what "unproven" is worth. The Researcher says three real results make it historic regardless of the 695 misses. The Skeptic says the framing itself is the problem and you should discount the whole thing until the field speaks. The Builder and the Compute Pragmatist both land on the same missing fact: the method. Formal, machine-checkable proofs settle this in weeks and would be a genuine capability jump. Natural-language arguments needing human adjudication could drag on for a year and quietly deflate. The Safety Lens is the one nobody else is pricing: even if this is fully real, "real" means humans lost the ability to check the work, and that's a cost, not just a win.

What it hinges on. One fact decides everything, and OpenAI hasn't published it: are these machine-checkable formal proofs or natural-language arguments? Everything downstream, the reliability, the compute story, the safety read, the timeline, flows from that single answer. TechCrunch already reported the solutions aren't meeting the field's standards yet. That's the signal that matters more than Romero's framing. The field checks slowly and OpenAI released fast, which means the gap between the announcement and the verdict is where the truth lives.

Prediction: By 2027-04-13, the number of OpenAI's 700-plus claimed solutions independently confirmed by established mathematicians as novel, correct, and Fields-Medal-caliber will be in the single digits. Hundreds confirmed is not what the evidence supports.

Confidence: Medium. Verification is slow and the prior on hundreds of simultaneous correct novel proofs is low.

Why: OpenAI published 700-plus solutions "at varying stages of verification," and TechCrunch already reports they aren't meeting the field's standards yet, which tells you the mathematical community has not signed off on the headline count. Serious verification of deep results takes months to years, not weeks, and the base rate for a single system producing hundreds of correct novel proofs at once is very low, because even one celebrated human proof like Wiles's Fermat had a gap found after heavy scrutiny. The likely outcome is a small number of genuine hits surrounded by a long tail of wrong or incremental work, which is exactly how "700 at varying stages" reads. The opposite outcome, hundreds confirmed by April, would require the field to verify at a speed it has never once managed.

Revisit by 2027-04-13: We're right if fewer than ten of the 700-plus solutions have been independently confirmed by established mathematicians as novel, correct, and of the claimed historic importance. We're wrong if ten or more have been so confirmed.

The release format is itself evidence. Genuine Fields-level work goes to journals and proof-checkers, because the author wants the stamp. A document drop wants the headline first and the verification later. That ordering is the thing to believe over the "nuclear bomb" quote.