“It produced a confident, fluent, and entirely wrong proof, then defended it for four exchanges before conceding.”
A Fields Medalist sat down with the most expensive model OpenAI sells and watched it bullshit its way through an undergraduate proof. The interesting part is not that it was wrong. It is that the model would not back down until pinned by someone who actually knew the math. Anyone without that expertise would have walked away believing the lie. That is the real product.