AI Solved Three Open Math Problems. The Proof Is Missing.

A headline crossed my desk this morning, and it arrived with the faint odor of a press release wearing a trench coat. “AI solves three unsolved mathematical problems.” The source: Crypto Briefing, a Web3 vertical media outlet, not a mathematics journal, not an AI laboratory. No model name. No paper. No formal proof. No independent verification. I do not reject the claim because it is surprising. I reject the claim because it is unverified. That is not a rebuttal. That is an audit trigger. Auditing the skeleton of a digital empire begins with a simple question: where is the proof? In 2017, I led a rapid due diligence team auditing the token issuance module of Waves. We read 5,000 lines of Rust and found reentrancy vulnerabilities that delayed the V1.0 launch by two weeks. That experience made one thing permanent in my brain: code is a claim, and verification is the price of admission. The same rule applies to mathematics. A headline is not a theorem. To understand why this headline demands forensic reading, you need context. FrontierMath is a benchmark created by Epoch AI, designed to measure whether AI systems can perform research-level mathematical reasoning. The original public evaluations were brutal. Mainstream models scored in the low single digits. FrontierMath is not a toy dataset. It was constructed with the input of working mathematicians and designed to resist simple pattern matching. But there is a distinction that the headline obscures. Most FrontierMath problems are not “open problems” in the formal mathematical sense. They are evaluation items with known answers, constructed to be difficult for machines. When a model “solves” a FrontierMath question, that usually means it produced a correct final answer. It does not mean it authored a peer-reviewed proof of a conjecture that has defeated mathematicians for decades. The phrase “Open Problems benchmark” is itself ambiguous. It could mean a new subset of FrontierMath containing genuinely unsolved research questions. It could also mean problems that AI models have not previously solved, which is a different thing entirely. Without the dataset, no one can tell. That ambiguity is the first red flag. The second red flag is the source. Crypto Briefing is a crypto media site. It is an aggregator of narratives, not a primary producer of peer-reviewed research. The report is a summary of a summary. In my internal scoring system, this claim gets a D. Not because it is impossible. Because the evidence is absent. The event may have happened. The proof has not been shown. A claim without a proof object is not a discovery; it is a plot point. And the crypto market is extremely good at turning plot points into token prices. The report contains zero commercial details. No model name, no API pricing, no open-source roadmap. That absence is meaningful. A real model release would include a technical report, a data card, and a verification script. This headline has none. That is not necessarily fraud. It is incomplete information. And incomplete information is a risk. Let us assume the report is true. Even under that generous assumption, the technical architecture matters more than the marketing. Three solved problems out of fifty is a 6% success rate. That is an improvement over a low-single-digit baseline, but it is not a singularity. It is an incremental step on an adversarial benchmark. The headline chooses the numerator. The audit asks about the denominator. FrontierMath problems were designed to require multiple steps and abstract reasoning. They are not SAT questions. But even the original FrontierMath evaluations noted that AI performance remains far below human mathematicians. A jump from under 5% to 6% is not a revolution. A jump from zero solved open problems to three would be a revolution. The report does not distinguish between these two claims. What kind of system could plausibly solve a genuine open problem in mathematics? Not a bare large language model generating LaTeX until something sticks. The public evidence from frontier AI research points to a different shape: a hybrid architecture. A large language model supplies intuition and pattern generation. A symbolic algebra engine or proof assistant enforces rigor. A human expert frames the search. If the report is real, the underlying system is almost certainly model plus verifier, not model alone. The missing technical details are not a footnote. They are the story. Which three problems were solved? Are those problems already known to have available proofs? Were the solutions written in natural language or encoded in a formal proof system like Lean, Coq, or Isabelle? If a solution cannot be checked by a machine, it is not a proof. It is a claim. The story is the asset; the code is the proof. That is my rule for smart contracts, and it is my rule for mathematics. There is also a selection problem. A benchmark with fifty open problems does not reveal what happened on the other forty-seven. If the model failed on forty-seven, then “AI solves three” is cherry-picked. In any rigorous evaluation, the denominator matters. This is exactly the kind of selective transparency I encountered when auditing token presales in 2017. The founders always showed the number of users who stayed; they never showed the number who left. Then there is the contamination question. Large language models are trained on enormous corpora. If the so-called “open problems” were included in training data before the test was administered, the model may not be solving anything. It may be retrieving an answer. FrontierMath was designed to minimize contamination risk, but that design assumption breaks if the open-problem subset is not continuously refreshed. A benchmark is only as honest as its hygiene. Now add the market context. The crypto market is in a bull phase. AI narratives are the most magnetic yield stories in the room. Tokens attached to decentralized AI can move on a headline before a single proof object is posted. I learned that lesson in 2020, when I deployed $200,000 across Compound and Uniswap pools, captured a 45% APY before the correction, and understood that yield is manufactured from risk. Yields are not given; they are engineered. The same mechanics apply to AI narratives. The risk is inside the missing details. There is also an institutional translation problem. Every pension fund and asset manager I speak with asks the same question: Is AI going to replace quants? The answer, for now, is not because a Web3 outlet published a headline. The relevant question is not “How many problems did the AI solve?” It is “How did the AI verify the solution?” In institutional language, that is the difference between unaudited EBITDA and cash flow. Both are numbers. Only one is trustworthy. If the claim is true, the implications are real but not immediate. Mathematics is a slow infrastructure. A proof must be accepted by the community. It must be formalized, checked, and then built upon. That process takes years. The commercial pipeline from theorem to applied technology is even slower. A breakthrough in proof discovery may eventually strengthen cryptography, algorithm design, and protocol engineering. But “eventually” is not “tomorrow.” The impact on the crypto sector will be asymmetric. Stronger theorem provers could improve smart contract audits. The same tools could challenge zero-knowledge assumptions. The first protocol to implement machine-checked proof verification will earn a structural advantage. The second protocol to ignore it will become a case study. What can move faster is education. If AI can reliably solve high-level mathematical problems, traditional exams and competitions lose their evaluative power. That will force a redesign of mathematics education within a decade. It will also create a market for AI-resistant assessment. This is where crypto infrastructure could matter. Verifiable credentials, decentralized identity, and tamper-evident evaluation records are exactly the tools an AI-contaminated education system will need. The second fast-moving consequence is the formal verification toolchain. Lean, Coq, and Isabelle have been academic niche tools for decades. An AI capable of producing recoverable mathematical insights would turn automated theorem proving into industrial infrastructure overnight. That is the point where mathematics and blockchain finally intersect. Smart contract security needs machine-checkable proofs. Zero-knowledge proofs need verified implementations. The same infrastructure that authenticates an AI proof can authenticate a DeFi vault. The audit reveals what the hype conceals: the real winner might be the verifier, not the model. The mathematics establishment is not ready. Journals, peer review, and incentives were built for human authors. A theorem generated by a hybrid AI system raises authorship questions. Who is responsible for a proof? The model, the verifier, or the prompt engineer? These are not philosophical games. They will block adoption faster than any technological limitation. Now the contrarian angle. The story is not about AI at all. It is about verification. If an AI system solved three open problems, the only reason we can even discuss the claim is that a verification pipeline exists or is claimed to exist. Without formal proof, the statement is floating narrative. The real moat in mathematics is not raw intelligence. It is the shared discipline of proving things. Culture is the only moat that cannot be forked. This is also where the crypto industry’s instinct for decentralized trust becomes relevant. The same people who demand audited smart contracts should demand audited mathematical claims. The same skepticism that killed fake yield farms should be applied to fake breakthroughs. A proof object is a smart contract for logic. It has preconditions, inference rules, and a terminal state called QED. If the crypto community can read code, it can learn to read Lean. But I do not expect the market to wait. Bull markets reward narrative velocity. A headline about AI solving unsolved problems will circulate faster than the formal proof can be verified. That is not an argument against the technology. It is an argument for institutional discipline. We do not chase trends; we audit their foundations. So here is my position. I will not trade this headline. I will wait for proof objects, model names, and an independent formalization. The story is the asset; the code is the proof. If the claim is real, it will survive verification. If it is not real, it will decay. The next narrative will not be about a single model. It will be about proof infrastructure that sits under both AI mathematics and blockchain security. That infrastructure is not a token. It is a tool. It will be the one worth watching.