“These language programs do write some “new” things—they’re called “hallucinations,” but they could also be described as lies. Similar to how autocorrect is ducking terrible at getting single letters right, these models mess up entire sentences and paragraphs.”
“Lies” assumes an intelligence; subterfuge. It is anthropomorphizing software and its bugs.
The proper term is bug. When I write code and it delivers the incorrect response, that is a bug.
We need to stop with incorrect and factually flawed terms — ChatGPT doesn’t have hallucinations, it has incorrect and erroneous responses.
Bard for example, couldn’t nail how many indictments there were under the Clinton Administration. How many were convicted? Throw of the dice seems to be Bards way of solving it.
These aren’t lies, they are bugs. They are fundamental to the FAILURE of the system as a whole.
These bugs would be inexcusable in any other REAL software industry.
@feloneouscat "Throwing the dice" is actually not an entirely inaccurate high level depiction of how generative transformers work, as they're statistical language models.
FWIW, and this isn't at all me excusing the wildly overstated claims of LLM accuracy and capability, "hallucination" is a term of art in this field referring to the model creating assertions that are both false and not in the original data set. They're a big problem! LLMs lacking a concept of truth is inherent in their design.
@lenaoflune "’Throwing the dice’ is actually not an entirely inaccurate high level depiction of how generative transformers work”
I know. That’s why I said it.
“Hallucination” is a term that is wildly inaccurate, highly anthropomorphic, and really doesn’t describe the nature of the problem or indicate how the issue may be resolved.
A bug by any other name still goes on the punch list.