I'm not an expert on LLM's or on ML but I think that's not the way it works, it doesn't search sources, read them and write the answer like a human would do...

I'm not an expert on LLM's or on ML but I think that's not the way it works, it doesn't search sources, read them and write the answer like a human would do... I believe the way it works is it processes millions of documents, tokenizes them, that is, it splits and slices them into tokens, and then since it was millions, billions or trillions of documents, it can predict the most statistically likely token to follow given a context, iterating on that it can generate full text.