“Current large language models can achieve a 95% score on the GSM-8K grade-school math benchmark.”