“Caught in their own GPTards trap.”
That’s how Yann LeCun, Meta’s Chief AI Scientist, characterized the backlash after OpenAI researchers celebrated what they claimed were major mathematical advances with GPT-5.
Demis Hassabis, CEO of Google DeepMind, also commented, “this is quite embarrassing.”
According to The Decoder, OpenAI VP Kevin Weil claimed in a tweet that was later removed that “GPT-5 discovered solutions to 10 (!) previously unsolved Erdős problems and made headway on 11 more.” (Erdős problems refer to well-known conjectures introduced by mathematician Paul Erdős.)
But Thomas Bloom, a mathematician who runs the Erdos Problems website, responded that Weil’s statement was “a significant misrepresentation.” Although these problems were marked as “open” on Bloom’s site, he clarified that this only indicates, “I personally do not know of any paper that resolves it.”
Put simply, it’s incorrect to assert that GPT-5 solved previously unsolved problems. Rather, Bloom explained, “GPT-5 located references to solutions for these problems that I was not previously aware of.”
Sebastien Bubeck, an OpenAI researcher who had also promoted GPT-5’s achievements, later admitted that “only solutions already present in the literature were identified,” but maintained that this was still impressive: “I understand just how challenging it is to search through academic literature.”
