The way I’ve explained it before is that it’s like the autocomplete on your phone. Your phone doesn’t know what you’re going to write, but it can predict that after word A, it is likelly word B will appear, so it suggests it. LLMs are just the same as that, but much more powerful and trained on the writing of thousands of people. The LLM predicts that after prompt X the most likelly set of characters to follow it is set Y. No comprehension required, just prediction based on previous data.