Batch of 722 math papers from OpenAI causes shock and distrust among experts
The company says results produced largely by a not-yet-released model could expand the frontiers of knowledge; several texts have already been withdrawn or altered.
Translated from the Portuguese original by the Panorama newsroom. Read in Portuguese
Summary written by the Panorama newsroom from the sources listed below. Facts are attributed to the original publications.
In early October, OpenAI published 722 papers addressing 372 open math problems, produced largely by an as-yet-unreleased AI model. The analysis comes from Professor Melissa Lee of Monash University (Australia), in a text published by Folha, which describes the commotion in the mathematical community.
Some of the works claim advances on famous problems, such as the Riemann hypothesis and the Birch and Swinnerton-Dyer conjecture, both Millennium Prize problems, which pays US$ 1 million to whoever solves them. Since the release, several papers have been withdrawn or altered.
The author questions the method. OpenAI says it wants to promote progress in mathematics, but, according to her, a senior colleague found one of the texts so incomprehensible that he would reject it if he were a journal editor. ChatGPT itself, when she consulted it, classified at least one of the results as a serious hallucination.
The text also recalls that, in September, the company released a controversial solution to a case of the Navier-Stokes problem. Mathematicians Tristan Buckmaster and Levent Alpöge accused OpenAI of using their ideas and data, which the company denies.
For the author, another possible motivation is the race for artificial general intelligence, since pure mathematics is a symbolic terrain for showing reasoning ability.