OpenAI released hundreds of claimed solutions to complex math problems this week, saying it consulted mathematicians to avoid past controversies. The lab followed some AGMAI principles, including releasing results quickly and including model reasoning, but not for all 719 manuscripts.
The Advisory Group on Mathematics and Artificial Intelligence (AGMAI) urged OpenAI to stop testing advanced math problems on proprietary models. OpenAI said it evaluates its models using open research problems, but the advisory group did not provide a detailed evaluation of the latest proof release.
"Problems are being solved autonomously by AI prompters who have no interest in the broader field itself once their initial target is 'solved'," said Terence Tao, a mathematician who criticized OpenAI’s approach. He added that AI models often fail to engage with the broader mathematical community after solving a problem.
A paper by Cambridge and King’s College mathematicians highlighted discrepancies between natural language proofs and Lean code used to formalize solutions. These differences raise questions about whether AI can reliably formalize its own results without human oversight.
AGMAI recommended OpenAI include machine-readable metadata linking natural language and formal proofs, which the lab did not do. The authors of the 'lost in translation' paper argue that without peer review, AI-generated proofs should not be trusted.
Mathematicians stress that human understanding is crucial for new results, as it allows for application in practical fields and further research. OpenAI did not say whether it will adopt the AGMAI recommendations, leaving uncertainty about its next steps.
Source: techcrunch