An OpenAI researcher posted that 'GPT-5 has solved an unsolved mathematical problem,' but it turned out that the problem had already been solved, leading to ridicule from rival developers, including Google DeepMind CEO Demis Hassabis.



OpenAI has posted that it has solved several unsolved mathematical problems, ' Erdős' problems, ' posed by Paul Erdős , known as 'the mathematician who wrote the most papers in the 20th century.' However, it has become clear that it has simply searched through the literature for solutions to problems that have already been solved, leading to ridicule from rival AI developers.

Leading OpenAI researcher announced a GPT-5 math breakthrough that never happened
https://the-decoder.com/leading-openai-researcher-announced-a-gpt-5-math-breakthrough-that-never-happened/



OpenAI's 'embarrassing' math | TechCrunch
https://techcrunch.com/2025/10/19/openais-embarrassing-math/

On October 18, 2025, Kevin Weil, Chief Product Officer at OpenAI, posted, ' GPT-5 has solved 10 previously unsolved Erdős problems! It has also made progress on 11 other problems.' Weil claims that the problems solved by GPT-5 had been unsolved for decades. At the time of writing, Weil's post has been deleted.




Echoing Weyl, other researchers at OpenAI have also touted GPT-5's ability to solve unsolved mathematical problems. 'GPT-5 Pro is superhuman at literature search,' OpenAI computer scientist Sebastian Bubeck wrote in a post. 'It solved the Erdős problem, listed as unsolved in the official database , by realizing that it had already been solved 20 years ago.'




In response to Bubeck's post, Mark Selke, Associate Professor of Statistics at Harvard University, said, 'Using thousands of queries from GPT-5, we have found solutions to 10 Erdős problems that were listed as unsolved. The 10 we solved are 223, 339, 494, 515, 621, 822, 883 (Part 2/2), 903, 1043, and 1079. In addition, for 11 other problems (32, 167, 188, 750, 788, 811, 827, 829, 1017, 1011, and 1041), GPT-5 found significant partial progress that we added to the official website. Regarding 827, there was actually an error in Erdős's original paper, and Martinez and Roldán Pensado's research explains this and corrects the argument. The future of scientific research will be exciting.'




In response, OpenAI's Boris Power said, 'Wow, we've finally made some major progress on a previously unsolvable problem!'




However, mathematician Thomas Bloom, who runs the official Erdős Problems database, quickly refuted the OpenAI researchers' post, explaining that the problems listed on the site as unsolved were problems he personally didn't know how to solve, and were not actually unsolved.




In response, Demis Hassabis, CEO of Google DeepMind, a competitor to OpenAI, said, 'This is embarrassing.'




After Bloom's post, Bubeck wrote, 'I have deleted my original post. I obviously did not intend to mislead anyone. I thought my wording was clear, and I apologize. I only found solutions in the literature. That's all. And knowing how difficult it is to research the literature, this feels very accelerated.' He apologized for posting as if GPT-5 had independently generated proofs for difficult mathematical problems, and deleted the post.




In response to Bubeck's apology, Yann LeCun, Meta's chief AI scientist, wrote that he had been 'caught up in his own GPTards' (a slang term used to mock those who espouse excessive belief in AI tools such as ChatGPT).




Related Posts:

in Software, Posted by logu_ii