LLMs have already discovered new proofs for math problems that were previously unsolved. Granted, this hasn't been done with a commercially available model as far a I know, but you are technically wrong to say they will never discover anything new.