When AI Solves the Unsolvable, Check the Proof, Not the Press Release

In the past seventy-two hours, a benchmark called FrontierMath has become the most contested square of code in the artificial intelligence world. A report from Crypto Briefing claims an AI system has solved three unsolved mathematics problems from a new “Open Problems” subset. Three problems. One model. Zero released proofs. The headline should make any careful reader pause. Not because progress is impossible, but because the distance between a claim and a verified theorem is exactly the distance between a whitepaper and a working protocol. I spent years auditing ICO whitepapers during the 2017 mania, and I learned that the most explosive narratives are often built on the least auditable foundations. This one feels familiar. FrontierMath was created by Epoch AI to measure whether models can do research-level mathematics, not schoolbook calculus. Early public testing showed mainstream large language models solving fewer than a few percent of its problems. The benchmark is deliberately engineered to resist pattern-matching memorization. Problems require constructing proofs, inventing intermediate objects, and reasoning over many steps. An “Open Problems” subset, if that is what the report refers to, raises the stakes further: these are questions that professional mathematicians have not yet answered. Yet the report contains no model name, no official blog post, no paper link, no formal verification file, and no benchmark card. That does not mean the event is false. It means the report is a teaser, not a technical announcement. Assume, for a moment, that the claim is true. What is the plausible route? Based on public knowledge of FrontierMath, the most credible path is not a single LLM dictating a final proof. It is a hybrid system: a language model generating candidate conjectures, a formal theorem prover such as Lean, Coq, or Isabelle checking each logical step, and human experts guiding the search through the vast space of possible constructions. This is not exotic speculation; it is the architecture emerging across AI mathematics since 2024. I have seen the same hybrid pattern in security auditing. The best vulnerabilities are rarely found by a single scanner. They are found by a scanner generating hypotheses and a human verifying the exploit path. A proof is no different. Generation is cheap; verification is expensive. Now, the information gain that the original report omitted. If there are fifty problems in the benchmark and the model solved three, then it failed on forty-seven. That failure distribution is not a footnote; it is the primary dataset. Which three problems were solved? Were they the shortest? The ones with the most accessible formalization? Did the model solve them from scratch, or did it receive a hint about which theorem to use? Without that context, “three solved problems” is statistically meaningless. A model that cracks one genuinely difficult problem—one that resists the entire field for years—is a different story from a model that cherry-picks the easiest three from a curated list. The publication of the 47 failures would tell us more about the model's true capability than the three celebratory headlines. Verification matters even more. An unsolved math problem is not a multiple-choice question. A solution has to be a rigorous argument, not a numeric answer. If the AI's answer is in natural language, the mathematical community will need months, perhaps years, to verify it. If the answer is encoded in Lean, the theorem prover becomes the judge. Then anyone with compute can replay the proof and confirm its validity. That distinction is everything. Code doesn't care about reputation. Code doesn't get swayed by charisma. A formal proof is a kind of cryptographic hash: immutable, auditable, unforgiving. Trust must be engineered, not promised. This brings me to the contrarian angle. The real story is not that AI might be solving open mathematics. It is that humanity's bottleneck has never been generating proofs; it has been trusting them. The old system worked like this: a brilliant mathematician publishes a proof, and the community spends years checking it. Sometimes the proof is wrong; sometimes it is incomplete; sometimes it is correct for the wrong reason. Now, with AI, a model generates a proof in hours, and the verifier must be faster and more rigorous than the generator. That inversion changes the entire economy of mathematical research. The scarce resource is no longer insight. It is verification. We have lived this inversion in crypto. A protocol can claim a billion dollars in total value locked, and without auditable code, it is just a narrative. Soulless finance is just empty pixels. The same is now true for mathematics. An AI that solves three open problems without a formal proof trail is not a breakthrough; it is an advertisement. It occupies the same category as an unaudited smart contract: possible, promising, and not yet real. The solution must be replayed by an independent verifier, ideally in formal logic, before the mathematical community treats it as knowledge. Otherwise, we are not doing math; we are doing marketing. There is also the question of purpose. The report frames this as a story of capability. But the deeper issue is one of accountability. If a commercial AI lab solved these problems, why is the model still unnamed? Why did the report come through a crypto media outlet rather than a mathematics preprint server? There is a pattern in crypto: a team announces a breakthrough, the token pumps, and the technical details arrive later, if at all. With math, the technical details are the entire product. A theorem with no proof is not a theorem; it is a rumor. In my years of work, I have learned that truth requires human skin in the game. That phrase is not sentimentality; it is architecture. A zero-knowledge proof can show that a computation happened without revealing the computation. But someone has to define the statement being proved. Someone has to encode the question correctly. Someone has to challenge the assumptions. If an AI solves a math problem and no human reviews the formalization, then the “solution” may have solved a different problem than the one we think: a subtle mismatch in definitions, a missing axiom, or a false framing of the open question. Human verification is not a nice-to-have; it is the bridge between computation and knowledge. The next narrative is not “AI solves math.” It is “Who verifies the verifier?” The answer will be built in Lean and Coq, encoded in zero-knowledge proofs, and anchored by human reviewers who still understand that a proof generated without human accountability is just another token. We need a protocol for trust. Code doesn't end the debate; it begins it. When an AI claims to solve the unsolvable, the question is not whether the model is brilliant. The question is whether the proof can be checked by a stranger in a different country with a different bias and a different incentive. If it can, we have progress. If it cannot, we have a press release. I know which one I would rather audit.