On July 20, 2026, the mathematical world was rocked by claims that OpenAI’s GPT-5.6 Sol, working with mathematician “Korsky,” had solved a decades-old problem from the legendary Paul Erdős—and in just one page. The news spread rapidly through tech and science outlets, but a check of the official Erdős Problems database that same day told a different story: Problem #119 remained stubbornly listed as “OPEN,” with its $100 prize unclaimed. So, what really happened?
The Claim: A One-Page Wonder
According to a report from KuCoin, GPT-5.6 Sol produced a proof that eclipsed a 44-page paper published by József Beck in 1991. Thomas Bloom, the University of Manchester mathematician who curates the Erdős Problems website (erdosproblems.com), praised the new argument as “stronger than Beck's” and constructed using only “elementary harmonic analysis techniques.” The key maneuver involved convolving a non-negative function, shifting it in time, and then estimating an operator norm through duality—a clever but not unprecedented trick in analysis.
The proof was a collaboration: a mathematician identified as Korsky guided the AI, which generated the one-page argument. Bloom’s initial public comments on X (formerly Twitter) fueled excitement, with OpenAI President Greg Brockman calling it “a watershed moment for mathematics.” Yet, for all the fanfare, the official problem database hadn’t budged.
A Problem with Three Parts
To understand the gap between the headline and reality, you have to know what Erdős Problem #119 actually asks. It’s not one question but three, all about the maximum modulus of polynomials with zeros on the unit circle. Over the years, two of the three parts were solved—but the third, the one with the $100 prize, had remained open.
| Question | What It Asks | Status (Before July 20) |
|---|---|---|
| First | Does the maximum modulus M_n grow without bound along some subsequence? | Solved by Wagner (1980) |
| Second | Is there a constant c>0 such that M_n > n^c for infinitely many n? | Solved by Beck (1991) |
| Third | For some c>0 and all large n, does the cumulative sum ∑_{k≤n} M_k exceed n^{1+c}? | Open, $100 prize unclaimed |
Beck’s 44-page paper settled the second question, not the third. The cumulative lower bound demanded by the third question is significantly stronger: it requires persistent, power-like growth of the average, not just occasional spikes. So when headlines screamed that AI had solved “Erdős’s toughest problem,” they were conflating the second and third parts.
What the One-Pager Might Have Done
The public report does not include the actual manuscript, so its precise theorem remains unknown. Bloom’s phrasing—“stronger than Beck’s”—could mean the proof establishes the third question, or it could simply mean it proves Beck’s second-question result with a better constant or cleaner method. Until the document is released, we can’t tell whether the $100 prize has truly been earned or whether the AI merely found an elegant shortcut for an already-solved piece.
The distinction matters. A one-page simplification of Beck’s work would be a beautiful contribution—it could make the theorem easier to teach, generalize, and formalize—but it wouldn’t close the book on Problem #119. Only a correct proof of the cumulative sum inequality would do that, and on July 20, no such proof had been publicly vetted.
The Verification Gap
Mathematics has no central authority that instantly certifies a proof. Verification emerges through a distributed process where specialists scrutinize every step, compare with existing literature, and try to break the argument. An AI can produce polished technical prose far faster than experts can audit it, and fluency is not evidence of correctness.
Bloom’s role is to maintain the database, not to confer final approval. He has pointed out in the past that a problem’s “open” status merely reflects his personal knowledge of the literature; newly discovered solutions often take time to incorporate. The website’s last update for Problem #119 was on January 23, 2026, so a recent submission might simply be under review. Broader confirmation would likely require a publicly accessible manuscript, scrutiny from multiple analysts, and ideally a formal verification using a proof assistant like Lean or Coq.
The history of AI mathematics claims is littered with premature announcements. In 2025, OpenAI touted solutions to several Erdős problems that turned out to be literature rediscoveries—the model had found existing papers, not generated new proofs. That incident prompted a sharp correction from Bloom and skepticism from figures like Yann LeCun. The current claim may prove different, but the cautious approach is to wait, not to declare victory based on a tweet.
Why Windows Users Should Care
An abstract theorem about complex polynomials might seem remote from your PC, but the technology behind it is rapidly coming to desktops. GPT-5.6 Sol is the flagship of a model family that powers productivity tools, code assistants, and research environments—many of which run on Windows.
- Copilot and beyond. Microsoft’s Copilot in Office, Visual Studio, and Windows itself is evolving from simple autocomplete to multi-step reasoning. The same persistence that let GPT-5.6 explore 64 parallel agents on a math problem can soon help you debug a complex application or analyze a spreadsheet.
- The need for provenance. If your research assistant begins generating conjectures or proofs, you’ll want an inspectable trail. Which sources did it consult? Which sub-agents generated each claim? Was the result formally checked? Expect future Windows AI tools to include logs, sandboxed execution, and integration with proof assistants and symbolic computation engines.
- Formal verification goes mainstream. Tools like Lean and Coq have been the domain of specialists, but AI can translate natural-language math into formal code and fill routine proof steps. A Windows user might one day ask an assistant to explain a theorem, generate a proof, formalize it, and run the checker—all from a familiar desktop interface.
- Caution is still king. The same fluency that impresses in mathematics can generate convincing nonsense in other domains. Always cross-check AI-generated analytical work with established tools and, where possible, demand machine-checked verification.
The Patience Factor: How AI Thinks Differently
Human intuition is an extraordinary compression mechanism—it lets experts discard thousands of dead ends without trying them. But occasionally, the discarded path contains a brilliant shortcut. An AI has no embarrassment, no career risk, and no boredom. It can try subtle variations, abandon them, and circle back, over and over. That’s how a one-page proof using convolution and a shift might finally surface after decades: not because the math was new, but because no human had the patience to keep tweaking parameters.
This shift from “solve” to “search exhaustively” is what makes multi-agent reasoning so powerful. By running 64 sub-agents in parallel, GPT-5.6 can simultaneously explore literature, test finite cases, critique candidate arguments, and synthesize the best result—turning a lone chatbot into a virtual research workshop. The innovation is as much about orchestration as it is about raw mathematical talent.
What You Can Do Now
For Windows users watching this story unfold, here are practical steps:
- Follow the official source. Bookmark erdosproblems.com and monitor the entry for Problem #119. When the database status changes or a preprint appears, you’ll get the most reliable signal.
- If you use AI for math or code, verify. Tools like Copilot can accelerate work, but treat their outputs as drafts. For critical analysis, use formal methods when possible, and always sanity-check with domain experts.
- Demand transparency. When AI contributes to a published result, ask: Who designed the prompt? What tooling was used? Was the proof independently reviewed? Reputable outlets will provide these details.
- Experiment with formal proof assistants. Lean, for example, has a growing community and can run on a standard Windows machine. As AI improves at generating formal proofs, this skill will become increasingly valuable.
Outlook: Beyond the Headlines
The alleged one-page Erdős proof may eventually be validated, classified as a simplification of Beck’s work, or rejected after peer review. On July 20, 2026, the public record was insufficient to choose among those outcomes. Yet the episode marks a genuine shift: AI is producing arguments serious enough that experts must inspect them, while the rest of us are learning that a proof headline is not a proof.
The next months will likely bring the manuscript’s release, formal verification attempts, and—one hopes—a definitive update to the Erdős Problems database. In the meantime, the lesson for Windows users is clear: powerful reasoning tools are on your desktop, but they’re only as trustworthy as the verification you wrap around them.