OpenAI’s math solutions aren’t meeting the field’s standards yet
techcrunch ·
OpenAI recently released hundreds of solutions to complex mathematical problems, claiming consultation with an advisory group to ensure high standards. However, the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) indicated that OpenAI fell short of these expectations, particularly regarding human understanding of the results. While OpenAI followed some guidelines by releasing proofs quickly, it failed to meet others, such as providing chain-of-thought data for most manuscripts or formalizing a significant portion of the proofs. Critics, including prominent mathematician Terence Tao, argue that AI solvers lack genuine interest in the broader field and cannot adequately explain their outputs. A new study highlighted discrepancies between natural language explanations and formal Lean code in OpenAI's solutions to a million-dollar problem derived from Navier-Stokes equations. These mistranslations raise concerns about relying on models to formalize their own work without rigorous human peer review. Mathematicians emphasize that true discovery requires human engagement and responsibility, which current AI processes do not provide. Consequently, the community remains skeptical about whether OpenAI is taking sufficient responsibility for ensuring human comprehension of these autonomous solutions.
techcrunch