>It can't reason about new things

In my tests, LLMs can work with token sequences guaranteed to not be found anywhere (randomly generated words/names, for example) and give meaningful answers.