“After large language models achieved high performance on Francois Chollet's ARC v1 benchmark, their performance on the subsequent ARC v2 benchmark dropped to nearly 0%.”