OpenAI: 700 Mathematical Preprints, Criticism on Transparency

OpenAI has published over 700 preprints of mathematical results on GitHub. Mathematicians and researchers are voicing concerns about a lack of information, organizing around a letter signed by 28 Fields Medal winners, and demanding guarantees of transparency. The company mentions an advisory group at the Institute for Advanced Study but does not elaborate on its functioning.

Scientific Validation and the Scope of Results Remain Unclear

The GitHub deposit has left many mathematicians unanswered on essential points: it is not specified how many problems have actually been solved or by what methods. Gary Marcus, cognitive scientist and emeritus professor at New York University, believes that the publication would not pass peer review due to the lack of clarity regarding the procedure followed. According to him, the initial report allows for "almost nothing" to be concluded. OpenAI refers to this as a "first step" but does not detail the workings of its new advisory group or any potential sharing of its most advanced models. At this stage, it is not established whether the produced evidence will advance human understanding or add to the confusion.

Letter from the Fields and Critiques: Misaligned Objectives

Twenty-eight Fields Medal winners have launched the Math and AI website to accuse AI companies of pursuing goals that are "severely misaligned" with those of the discipline. The Fields Medal is awarded every four years, typically to three or four researchers, and is often presented as the equivalent of a Nobel Prize in mathematics. Curtis McMullen, a professor at Harvard and a signatory, criticizes the lack of coordination with specialists and the absence of applications that concretely benefit society. He denounces a deployment carried out without consideration for the disruption of the academic ecosystem, viewing it as a quest for financial valuation. Researchers also point to a lack of transparency regarding the attainment of results and a misunderstanding of the importance of the problems addressed, topics that OpenAI has not clarified since its announcement. McMullen reminds us that the central goal of mathematics is to enhance the understanding of abstract structures, which supports technological progress, and argues that "massive resolution" out of context diverges from this aim.

What OpenAI Announces: 700 Preprints and an Advisory Group at the IAS

OpenAI has published over 700 mathematical preprints on GitHub and claims that its internally developed models solve an unprecedented number of long-standing problems to "push the boundaries of human knowledge." The company also indicates the creation of an advisory group hosted at the Institute for Advanced Study. A spokesperson did not immediately respond to requests for comment. Scientific American describes a field "already in shock," placing this excitement a month after the announcement of the resolution of the Navier–Stokes equation, a problem associated with a $1 million prize.

Mixed Reactions and Rapid Trajectory of Models

Opinions vary. Levent Alpöge, a mathematician at Anthropic, praised the OpenAI team on X, while noting "sad stories" of users being outpaced, and described it as "the most significant moment in mathematical history." The pressure surrounding the impact of AI on mathematical research was already mounting before these announcements. The first major models, primarily trained on texts, made elementary mistakes and hallucinations, but then progressed rapidly, with some now labeling them as "geniuses," even though others believe they are far from replacing experts. More broadly, AI is entering the realm of mathematics with enhanced capabilities but without necessarily grasping the mindset of researchers, in a context where its effects are already being felt across many sectors and where the community is expressing its frustration. These systems have also ingested vast corpora of human knowledge.