Meta’s chief AI scientist, Yann LeCun, recently weighed in on the backlash against OpenAI’s claims regarding the math capabilities of its latest model, GPT-5. LeCun’s comment, “Hoisted by their own GPTards,” captures the sentiment following OpenAI’s announcement celebrating GPT-5’s alleged advancements in solving complex mathematical challenges.
Google DeepMind CEO Demis Hassabis expressed similar dismay, labeling the situation as “embarrassing.”
The Decoder reported that OpenAI VP Kevin Weil had previously tweeted about GPT-5’s purported success in solving ten previously unsolved Erdős problems and making progress on eleven others. These problems, named after the renowned mathematician Paul Erdős, have intrigued mathematicians for decades.
However, mathematician Thomas Bloom, curator of the Erdős Problems website, criticized Weil’s statement as a “dramatic misrepresentation.” Bloom clarified that while these problems were classified as “open” on his site, this designation merely indicates a lack of awareness of existing solutions by him personally, not an absence of solutions overall.
In a follow-up, OpenAI researcher Sebastien Bubeck admitted that GPT-5 merely located existing solutions within the literature, not new breakthroughs. Nonetheless, he emphasized the difficulty of literature searches as a significant achievement in its own right.
Key Points:
- Claims of Achievement: OpenAI’s GPT-5 was said to have solved unsolved Erdős problems.
- Criticism: Statements about these achievements have been labeled misleading by experts.
- Clarification: GPT-5 found existing references rather than presenting novel solutions.
This episode underscores the ongoing discourse surrounding AI capabilities and the importance of accuracy in scientific communications.
