“Grant Sanderson, citing Alex Kontorovich, argues that even a small error rate in AI-generated natural language math proofs would be "insufferable" for human mathematicians due to the high cost of verification.”