I really wish the models were smart enough to "tail their own logs" and tell us whose work/papers/chats were the catalyst for these math insights. What makes their "internal" model so much better at math? Must be trained on a whole lot of honesty.