‘Embarrassing’ math from OpenAI

“Taken down by their own GPT hype.”

That’s how Meta’s AI Chief, Yann LeCun, characterized the backlash after OpenAI researchers celebrated supposed breakthroughs by GPT-5 in mathematics.

Google DeepMind’s CEO, Demis Hassabis, commented that “this is embarrassing.”

The Decoder reports that OpenAI VP Kevin Weil stated in a now-deleted tweet that “GPT-5 discovered solutions for 10 (!) previously unresolved Erdős problems and advanced work on 11 others.” (“Erdős problems” are well-known conjectures from mathematician Paul Erdős.)

However, Thomas Bloom, who manages the Erdos Problems website, stated that Weil’s post was “a significant misrepresentation”—while these problems were marked as “open” on Bloom’s site, he clarified that it only meant, “I am personally unaware of a published solution.”

In other words, it’s inaccurate to state that GPT-5 could resolve previously unsolved problems. Instead, Bloom wrote, “GPT-5 located references, which contained solutions to these problems, that I was personally unaware of.”

Sebastien Bubeck, an OpenAI researcher who had also been promoting GPT-5’s achievements, then admitted that “only solutions already in the literature were found,” but he suggested this is still a genuine achievement: “I know how difficult it is to search the literature.”