OpenAI’s latest boast about GPT-5’s mathematical prowess has drawn sharp criticism from the AI community — and some public ridicule from leading researchers.
It started when OpenAI executives celebrated GPT-5’s supposed ability to crack “10 previously unsolved Erdős problems,” a reference to a collection of long-standing mathematical conjectures attributed to Paul Erdős. According to a now-deleted tweet from OpenAI VP Kevin Weil, the model had also “made progress on 11 others.” The tone was triumphalist, and the claim spread quickly — until mathematicians began checking the details.
Thomas Bloom, who curates the Erdős Problems website, quickly refuted the claim, calling it “a dramatic misrepresentation.” The issues GPT-5 had “solved” weren’t open at all — at least, not to the mathematical community. Bloom clarified that the problems were merely listed as open because he was unaware of existing solutions in the literature. GPT-5 hadn’t discovered new proofs; it had simply found existing papers that already contained them.
As Bloom put it in his response, cited in TechCrunch, “GPT-5 found references, which solved these problems, that I personally was unaware of.” In other words, GPT-5 performed an impressive literature search — not a mathematical breakthrough.
The backlash was swift. Meta’s Yann LeCun called the situation “embarrassing,” mocking OpenAI’s overreach with the phrase, “Hoisted by their own GPTards.” DeepMind’s Demis Hassabis echoed the sentiment, saying bluntly that the episode was “embarrassing.”
Even within OpenAI, there was some backpedaling. Researcher Sebastien Bubeck acknowledged that the model hadn’t produced new mathematics, though he defended the effort: “I know how hard it is to search the literature,” he said, suggesting that GPT-5’s retrieval ability still represents meaningful progress.
The incident captures a broader tension in the current AI race: the gap between genuine discovery and automated search. Large language models are increasingly good at navigating vast corpora of scientific literature, surfacing obscure connections and forgotten papers. But the temptation to frame these abilities as original insight remains strong — especially in a field driven by hype cycles and headline-grabbing milestones.
GPT-5 didn’t solve Erdős problems, but it did highlight a very human one — how easily enthusiasm for AI’s potential can outpace its actual achievements.