Replying to @wazoobonkerbrain@lemmy.world
It has to do with how LLMs work. Put much, much too simply, to arrive at a response, they use weighted vectors (a line that conveys magnitude and direction) to find the next token that best follows the one before it.