The "stolen chain of thought" paper revealed GPT models using words in a non-literal, code-like manner, similar to spy language