OpenAI’s ‘embarrassing’ math | TechCrunch

“Caught in their own hype.”

This aptly captures the sentiment expressed by Yann LeCun, Meta’s Chief AI Scientist, regarding the backlash that followed OpenAI’s claims about GPT-5’s mathematical advancements.

Demis Hassabis, CEO of Google DeepMind, weighed in, stating that the situation was “embarrassing.”

According to a report from Decoder, a now-deleted tweet from OpenAI’s VP, Kevin Weil, proclaimed that “GPT-5 found solutions to 10 (!) previously unsolved Erdős problems and made progress on 11 more.” The term “Erdős problems” refers to renowned mathematical conjectures introduced by the mathematician Paul Erdős.

However, mathematician Thomas Bloom, who curates the Erdős Problems website, described Weil’s claims as “a dramatic misrepresentation.” While these problems were indeed marked as “open” on Bloom’s site, he clarified that this merely indicates his personal lack of awareness of any solutions. In simpler terms, it’s misleading to assert that GPT-5 solved challenges that had stumped mathematicians. Instead, Bloom explained, “GPT-5 found references to solutions that I personally was unaware of.”

In response, OpenAI researcher Sebastien Bubeck, who had also been highlighting GPT-5’s achievements, acknowledged that “only solutions in the literature were found.” Despite this, he maintained that it still represented a significant accomplishment, noting, “I know how challenging it is to search the literature.”