OpenAI wants to consult elite mathematicians about how to not fumble again