> When producing full natural-language proofs on IMO-AG-30, however, GPT-4 has a success rate of 0% ...