Max LeibmanI don't like the term "hallucinations" when we talk about AI. Sure, LLMs can get things wrong, but a hallucination is an error in perception, and you can't have an error in perception when there's no one there to perceive. The only hallucinations that are happening are on your side of the keyboard.
Chris Smart@maxleibman That's a great point. What do we call them then? just "errors"?
@VE3RWJ That I don’t have a good answer to.
Corners Plotted@maxleibman @VE3RWJ Yes, it’s a (deliberately) difficult position!
I think part of the trickiness here is that the “hallucinations” aren’t materially different from what they do the rest of the time. It’s just that this response is so obviously wrong that we classify it as an error. But it’s not like something broke _that one time_. All responses are “hallucinations.” They vary by proximity to accuracy. The term is pure marketing.
Three plus or minus five@corners_plotted @maxleibman @VE3RWJ
It has some relationship to reality, a model that outputs false positives even though ground truth denies it; a bias to see patterns that don't exist.
But I agree with your assessment that it's not really something different than all the other output. It's just wrong. The AI makes EVERYTHING up, it's just that often it turns out to be similar to reality.
@ThreeSigma @corners_plotted @VE3RWJ Exactly. If you're guessing a statistically plausible next word, you're going to line up with reality often (maybe even shockingly often), because what's likely to come next will be something that makes sense, and making sense is often correlated with reality. But it's a correlation, not knowledge. There's no amount of grounding that makes it something other than a guess. Grounding is just the process of changing to the question to be, "Ok, given THIS context, what's NOW the most likely next word?"
@maxleibman @corners_plotted @VE3RWJ
I had someone try to convince me in another thread that LLMs didn't work word-to-word, but composed answers hierarchically in paragraphs or whatever.. My understanding is that that's wrong, and they work only on the next word, but maybe my understanding is a year or two out of date?