OpenAI’s ‘embarrassing’ math

Meta's Chief AI Scientist, Yann LeCun, characterized the backlash against OpenAI's recent claims about GPT-5 as being 'hoisted by their own GPTards.' This reaction followed a proclamation by OpenAI that their latest model had achieved remarkable breakthroughs in mathematics, specifically in solving Erdős problems. Demis Hassabis, CEO of Google DeepMind, joined the chorus of criticism, calling the situation 'embarrassing.' The controversy erupted after OpenAI's Vice President, Kevin Weil, made a now-deleted tweet boasting that GPT-5 had found solutions to ten previously unsolved Erdős problems, along with making progress on eleven others. Erdős problems are significant conjectures introduced by the renowned mathematician Paul Erdős. However, mathematician Thomas Bloom, who oversees the Erdős Problems website, contested Weil's claims. Bloom described the assertion as 'a dramatic misrepresentation,' clarifying that while these problems were marked as 'open' on his site, it merely indicated that he was unaware of any existing papers that addressed them. He emphasized that GPT-5 did not resolve these problems but rather located references to existing solutions that he had not previously encountered. In response to the unfolding debate, Sebastien Bubeck, an OpenAI researcher who had also celebrated GPT-5’s achievements, conceded that the model only identified solutions already documented in the literature. Nevertheless, he maintained that this task is a noteworthy accomplishment, stating, 'I know how hard it is to search the literature.' This exchange highlights the ongoing scrutiny and high expectations surrounding advancements in AI technology.