“The best models' accuracy on the MATH benchmark improved from approximately 5% to 50% within one year by mid-2022 and has since reached over 90%.”