OpenAI claimed a math breakthrough this weekend, only to be smacked down
The embarrassing episode sprouted from a misunderstood post, amplified by an OpenAI executive as proof of GPT-5’s mathematical prowess, but turned out not to be what it seemed.
In an embarrassing episode this weekend, OpenAI was forced to walk back a claim of a major math breakthrough. Kevin Weil, OpenAI’s VP of science, posted on X that “GPT-5 just found solutions to 10 (!) previously unsolved Erdős problems, and made progress on 11 others. These have all been open for decades.”
What do you think “found solutions” means? It sure sounds like they are saying GPT-5 solved these problems!
But shortly after posting this, Thomas Bloom, who runs ErdosProblems.com, disputed the claim. Bloom’s website contains a database of 1,103 of Hungarian mathematician Paul Erdős’ challenging problems and references to their solutions that have been found in academic research papers. Bloom called OpenAI’s announcement “a dramatic misrepresentation,” and said that GPT-5 found merely references to solutions for 10 problems that were previously published.
The competition is extremely intense among top AI companies to achieve computing breakthroughs that might indicate the eternally imminent arrival of AGI, so a post like this from such a high-ranking executive caught the industry’s attention.
After being called out for the incorrect claim, Weil acknowledged the error, saying he misunderstood the post in which researcher Mark Sellke detailed the Erdős problems listed as open that GPT-5 had found published solutions for. Weil deleted the original post, but not before other top AI researchers from OpenAI’s rival’s weighed in.
Alphabet’s GoogleMind cofounder and Nobel Prize winner Demis Hassabis wrote, “this is embarrassing.”
Meta Chief AI Scientist Yann LeCun, a pioneer in neural networks, dropped the mic with this post:
Hoisted by their own GPTards
— Yann LeCun (@ylecun) October 18, 2025