“For the ARC prize, training a model on millions of brute-forced, generated tasks to rely on memorization and overlap with the test set would be considered cheating.”