OpenAI's Math Struggles: An 'Embarrassing' Weakness?

OpenAI's Claims Regarding GPT-5 and Erdős Problems Face Scrutiny

A recent claim by OpenAI regarding the capabilities of GPT-5 has drawn criticism, with Meta’s chief AI scientist, Yann LeCun, stating the situation as being “hoisted by their own GPTards.” This comment followed OpenAI researchers’ initial celebration of alleged mathematical breakthroughs achieved by the model.

Demis Hassabis, CEO of Google DeepMind, further expressed his disapproval, labeling the incident as “embarrassing.”

Initial Claims of Solving Unsolved Problems

According to a report by The Decoder, OpenAI VP Kevin Weil initially announced via a now-deleted tweet that GPT-5 had successfully found solutions to ten previously unsolved problems formulated by mathematician Paul Erdős, and also made headway on eleven additional ones.

These “Erdős problems” represent a collection of famous, long-standing conjectures in the field of mathematics.

Dispute and Clarification

However, mathematician Thomas Bloom, the curator of the Erdos Problems website, refuted Weil’s assertion, characterizing it as “a dramatic misrepresentation.”

Bloom clarified that the “open” status of these problems on his website simply indicates his personal lack of awareness of any published solutions.

Essentially, GPT-5 did not independently solve previously unsolved mathematical problems. Instead, the model identified existing references containing solutions that were previously unknown to Bloom.

Acknowledged Limitations and Remaining Significance

Sebastien Bubeck, an OpenAI researcher who had previously promoted GPT-5’s achievements, subsequently conceded that the solutions discovered were already present in existing literature.

Despite this, Bubeck maintained that the accomplishment is noteworthy, emphasizing the difficulty of effectively searching through extensive academic literature. He stated, “I know how hard it is to search the literature.”

Key takeaway: While GPT-5 didn't create new solutions, its ability to locate existing ones demonstrates a capacity for advanced information retrieval.