OpenAI unreleased 4000-problem math model. Is this the first true look at GPT-7?

[Sevam]

OpenAI just dropped a massive collection of math research papers (722 papers across 372 families) generated by an unreleased internal model. This isn't a school test, the system was fed 4,000 open research problems that human experts haven't solved yet.
Each successful proof took an average of 3 hours of continuous reinforcement-learning compute. It handles everything from shapes to probability, with a fifth of the papers actually disproving long-standing assumptions.