OpenAI is in damage control mode after its brash announcement of its new artificial intelligence model GPT-5 and its power to compute complex mathematical equations. The initial triumphant announcement is now a humiliating backpedaling as some big names in both AI research and mathematics spoke critically on the matter.
The row began when Kevin Weil, Vice President of OpenAI, tweeted that “GPT-5 solved 10long-standing open Erdos problems and made significant progress on 11 others.” The use of exclamation marks and an effusive tone indicated that history had seen a dramatic breakthrough in mathematics research.
The problems are so-called after the great Hungarian mathematician Paul Erdos and are regarded as among the most difficult conjectures in mathematics. The tweet was later withdrawn.
The announcement promptly caused a stir within the AI community, although not quite as hoped for by OpenAI. Demis Hassabis, CEO of Google’s DeepMind and a respected name in artificial intelligence circles, was frankly critical. He referred to OpenAI’s endeavors as “embarrassing,” a frankly honest critique from one executive to another within the community.
GPT-5 of OpenAI Did Not Solve New Math Problems, Only Found Existing Solutions
The real blow came from mathematician Thomas Bloom, who maintains the Erdos Problems website, the very source OpenAI appeared to have consulted. Bloom was quick to set the record straight, calling Weil’s post “a dramatic misrepresentation” of what actually happened.
Bloom clarifies that when he tracks issues as “open” on his site, it doesn’t imply that those issues are necessarily open within the mathematical community.
It is simply that he himself hasn’t encountered nor was informed of a scholarly paper that would settle that specific issue.
The argument is rather significant. GPT-5 wasn’t revealing older unsuspected mathematical secrets; it simply encountered solutions already existing in academic literature that Bloom wasn’t listing on his site yet.
Fanning the flames still further was Meta Chief AI Scientist Yann LeCun, who is also a Turing Award laureate, with typically candid commentary. He framed the circumstances as OpenAI having been “hoisted on their own GPTards,” a sarcasm-laden reference to over-enthusiastic fans who might well have fed hype over the company’s boasts.
In its aftermath, Sebastien Bubeck, one of the OpenAI researchers who had been enthusiastically marketing GPT-5’s mathematical prowess, made a clarifying comment. He did admit that “only solutions in the literature were found,” in effect acknowledging that no new problems had actually been solved that weren’t already solved.
However, Bubeck attempted to salvage the narrative by arguing this still represents meaningful progress. “I know how hard it is to search the literature,” he said, suggesting that GPT-5’s ability to locate these references in academic papers demonstrates valuable search and comprehension capabilities.
What OpenAI’s Mathematical Misstep Teaches the Community?
This episode is symptomatic of a larger issue within the AI community: that fine line between legitimate enthusiasm over scientific advancement and false hype. With larger AI labs becoming increasingly competitive with one another, pressures for big announcements have never been greater.
The backlash also shows just how quickly technical accomplishments can be overhyped or misunderstood. Identifying solutions that are already available in research papers is certainly a good ability for a program like AI, but it is quite different from devising solutions to older mathematical challenges, something which did not become clearly explained with OpenAI’s earlier PR.
The community of mathematics was especially irritated with the incident. Real mathematical discoveries involve strenuous proof and peer examination, not mere AI-derived pronouncements that are later subject to excessive backtracking.
This isn’t just about one deleted tweet. It reflects growing concerns about how AI companies communicate their achievements to the public and investors. As these systems become more sophisticated, distinguishing between genuine breakthroughs and incremental improvements becomes increasingly important.
OpenAI’s gaffe is a lesson that in efforts to showcase AI excellence, accuracy and truthfulness should always be paramount. The initial announcement that created headlines and excitement was retracted later, which may even have discredited it more than would have the truth itself.
For the moment, those Erdos problems that truly are still open are still waiting for human mathematicians, or maybe someday for an AI program to solve them. But when that day arrives, the AI community should be much more reluctant to make its announcement.
