[pixl97]

LLMs don't see words like you do. Tokenization makes it behave odd. Often you can get 4 to output a solution in code and from that it derives a correct answer.