Re:Capital One to buy Discover Financial in $35.3B all-stock...
by NoPedantsThanks on Tuesday February 20, 2024 @12:10AM
(#39438270)
It doesn't "hallucinate." It provides incorrect output. That is all.
[ Parent
| on HN ]
Re:Capital One to buy Discover Financial in $35.3B all-stock...
by coder543 on Tuesday February 20, 2024 @12:14AM
(#39438284)
"Hallucination" is a widely understood and accepted term in the LLM industry. If you want to change that, replying to my comment doesn't seem to be the most effective place to start. I don't know that it's the best term, but it seems better than re-explaining the issue in detail every time.
"Incorrect output" is broader, encompassing other failure modes. If you ask an LLM to respond in JSON with a list of common foods, and it instead writes a paragraph of text that contains a list of common foods, then that would not qualify as a "hallucination" by my understanding of the accepted definition, but it is still "incorrect output".
Re:Capital One to buy Discover Financial in $35.3B all-stock...
by NoPedantsThanks on Wednesday February 21, 2024 @12:09AM
(#39450366)
"Incorrect output" is still correct to describe the cited behavior. "Hallucination" isn't. If a human behaves the same way when asked a question, we do not say he's hallucinating.
One way to devalue incorrect terminology is to call it out when you see it, and use something accurate instead. That's how people learn.