Not sure why someone downvoted you. That’s exactly what the term means in this context. It’s those confidently written answers that contain false or fabricated information.
And this seems like the biggest limitation for the LLM approach. The model just knows that a certain set of tokens tends to follow another set of tokens.
It has no understanding of what the tokens represent. So it does a great job of producing sentences that look meaningful, but any actual meaning in them is purely incidental.
Add comment