News for hackers. Stuff that matters.
Re:Weird GPT-4 behavior for the specific string “ davidjl” by JohnFen on Thursday June 08, 2023 @12:07PM (#36244204)
I also think "hallucination" is generally an actively terrible term for AI output that isn't factually correct. It's misleading, at the very least.
Re:Weird GPT-4 behavior for the specific string “ davidjl” by williamcotton on Thursday June 08, 2023 @12:09PM (#36244233)
I agree and unfortunately I didn’t construct my original comment in a manner that makes that agreement clear!
Re:Weird GPT-4 behavior for the specific string “ davidjl” by burnte on Thursday June 08, 2023 @12:43PM (#36244784)
I think it's incredibly accurate. Humans can misinterpret some internal inputs as external inputs and we hallucinate. This is doing the same, erroneous interpretation leading to nonsense interactions.
Re:Weird GPT-4 behavior for the specific string “ davidjl” by JohnFen on Thursday June 08, 2023 @12:47PM (#36244878)
I don't think that's an accurate description of these errors at all, honestly. It's not a matter of ChatGPT "interpreting" anything. It's a matter of it assembling a response that is linguistically most likely, given its training.

That's categorically different than hallucination, which is an active misinterpretation/invention of sensory data.

Re:Weird GPT-4 behavior for the specific string “ davidjl” by digging on Thursday June 08, 2023 @12:56PM (#36245034)
It "sees" something it "believes" is true, how is hallucination not an accurate term?
Re:Weird GPT-4 behavior for the specific string “ davidjl” by JohnFen on Thursday June 08, 2023 @01:00PM (#36245098)
Because it's neither "seeing" anything, nor is it "believing" anything.

In fact, I think it's incorrect to even call such responses "wrong". ChatGPT successfully put together a reasonable-sounding response to the prompt, which is exactly what it is designed to do.

Re:Weird GPT-4 behavior for the specific string “ davidjl” by williamcotton on Thursday June 08, 2023 @01:22PM (#36245443)
You're correct. For the downvoters, here is the explanation:

You know how sometimes an LLM completes a given text in a way that you would consider "wrong"? Well, the LLM has no concept of "correct" or "wrong". It just has a very broad and deep model of the entirety of the text in the training data. Sometimes you might consider the completion "correct". Sometimes you might consider the completion "wrong". Sometimes I may consider the completion "correct" and you may consider the completion "wrong", just as with many other kinds of statements. The logical reasoning happens outside of the statement itself and can reasonably have many interpretations.

Would you ever consider that a plant has grown in a way that is "wrong"? Or that a broken reciprocating saw cut through plywood in the "wrong" way?

Re:Weird GPT-4 behavior for the specific string “ davidjl” by digging on Thursday June 08, 2023 @03:15PM (#36247127)
Well, feel free to make up a new word then and see if it catches on. I will keep calling it hallucination because that effortlessly describes what is happening through analogy, a powerful tool by which we can make difficult concepts more accessible. I hope you realize that my use of quotes is another common literary tool which indicated I know a LLM can't actually see.
Re:Weird GPT-4 behavior for the specific string “ davidjl” by tiedieconderoga on Thursday June 08, 2023 @01:02PM (#36245136)
Because it neither "sees" nor "believes". It's a mathematical function that predicts the most likely next number in a series.
Re:Weird GPT-4 behavior for the specific string “ davidjl” by pegasus on Thursday June 08, 2023 @01:17PM (#36245365)
Agreed. I find "confabulating" more appropriate.