Do you disagree with my claim that GPT-4 can perform some sort of basic reasoning about puzzles that aren't in it's training data?