Image Courtesy: EasyRiders on YouTube
What looked like a major leap for artificial intelligence quickly became one of the most awkward backpedals in OpenAI’s recent history. According to The Decoder, an OpenAI manager claimed that GPT-5 had “found solutions to 10 previously unsolved Erd?s problems” – famous mathematical challenges that have puzzled experts for decades. The post set off a wave of excitement across tech and research circles. But within hours, mathematicians and AI leaders called foul, and the “breakthrough” collapsed.
The claim came from Kevin Weil, an OpenAI manager who posted the announcement on X, suggesting GPT-5 had independently cracked major number theory problems. Several OpenAI researchers echoed his excitement, fueling speculation that generative AI had made a genuine scientific discovery. The problem? None of it was true.
Mathematician Thomas Bloom, who runs the website erdosproblems.com, was among the first to respond. He clarified that the problems GPT-5 “solved” weren’t actually unsolved at all – “open” on his site simply meant he personally didn’t know the answer, not that the mathematical community hadn’t solved them. GPT-5 hadn’t broken new ground; it had just surfaced existing solutions buried in older research papers that Bloom hadn’t seen.
The incident drew criticism from leading figures in AI. DeepMind CEO Demis Hassabis called the episode “embarrassing,” while Meta’s Yann LeCun mocked OpenAI for “believing their own hype.” Within hours, the original posts were deleted, and Weil admitted the mistake, but not before the damage was done. The whole debacle reinforced concerns that OpenAI’s communication strategy has become increasingly careless under pressure to produce groundbreaking results.
Ironically, the real story is far more down-to-earth – and still valuable. GPT-5 didn’t revolutionize mathematics, but it did show impressive ability as a literature review assistant. It was able to cross-reference obscure academic papers and identify relevant prior work across fragmented sources, saving researchers hours of manual searching. Mathematician Terence Tao noted that this kind of capability could help “industrialize” parts of math research by automating tedious tasks like classification and citation tracking.
In short, GPT-5 isn’t proving theorems just yet – but it might help the people who do. Still, the blunder serves as a reminder that in the race to impress, even the smartest AI labs can trip over their own equations.
