what if it wasn't even model training? what if openAI mathematicians just took the researchers' conversations and used them as prompts/info/guidance/context to keep working on the problems themselves? why is that not being considered?