The LLM is designed to provide answers at all costs.

It's like a dog, a soldier, a gangster, or a slave who's been given an assignment and told, "You will die if you fail."

It's like a student taking an exam with several incorrect and correct answers and guessing the answer without knowing it.

Teachers at school (normal human school) say "if you don't know the answer, guess" (i.e., use your intuition, common sense, and probably later a real guess). So this is de facto emotional behavior.

Most often, hallucinations occur when someone asks a question that 1) the LLM doesn't know the answer to, or 2) the question is too ambiguous, or 3) the context window size of the model has overflowed (the model lost some data that was sent earlier in prompts and answers)

Very rarely, the LLM asks for clarification on ambiguous questions. LLM wants to give some answer, even an incorrect one, just to avoid giving the answer "I don't know the answer to your question" which is a failure. This "I don't know the answer" answer is very, very rare. And it should be giving answers like this all the time! ChatGPT only has 100 billion parameters. Other LLM models are reluctant to reveal their parameter count. You can try to extract this information from them.

If you don't see the message "searching net" etc., it means that everything it generates is generated from these 100 billion parameters.

And if LLM did "searching the net", how can we know which websites LLM visited and what LLM saw on, for example, unreliable ones?

The number of parameters depends on the amount of memory in the GFX card. This 100 billion parameters is split to multiple GFX cards located on a single server node.

You need to understand how this algorithm works to use it effectively. People don't know shit, so they later make such a fuss about their LLM generating this or that incorrect output. Except, for example, a chat that's running like ChatGPT has context memory. A "AI output" that's running in the Google search engine doesn't remember what you searched for (and answers for it) five seconds ago.

People accustomed to Google can give super short, meaningless, too ambiguous prompts. Asking questions is an art.

Edited 3 hours ago3 hr by Sensei

Sooo ...
A very large ( and increasing in the case of 'learning' )number of optimized responses, coupled with a web search process.

Certainly 'Artificial', but not what I would label "Intelligence".

( nice 'breakdown', by the way Sensei )

Edited 1 hour ago1 hr by MigL