>>509271886
>train ai on a bunch of math questions
Close, it's actually because of reinforcement learning on verifiable rewards, same way as AlphaGo etc. That's *on top* of the pretraining paradigm (training models 10,000x bigger than GPT4 before the end of the decade.)

>>509271992
There's a difference between "really good and improving rapidly" and "currently solving millennium problems." It'll definitely be solving open questions soon, though.