5 ms·
Hallucination is a terrible model for LLM factual errors, perhaps as terrible as I could think of. Most pharmacological true hallucinogens are all about the bra
by 65a 3y ago
Hallucination is a terrible model for LLM factual errors, perhaps as terrible as I could think of. Most pharmacological true hallucinogens are all about the brain not being able to properly receive input (anticholingergics, e.g. diphenhydramine, BZ, cyclizine, etc) and misinterpreting it with additional neuronal functional degradation due to synaptic interference by neurotransmitters (or lack thereof). This isn't really the problem that LLMs have, which in my opinion is much more akin to classical confabulation: https://en.wikipedia.org/wiki/Confabulation https://en.wikipedia.org/wiki/Confabulation
- kromem 3y agoWhile I don't think the term will switch at this point, you are dead right. From early on it reminded me of the anecdotes where people with a severed corpus callosum would have their left brain suddenly explain why they had just laughed, completely unaware their right brain has just been shown fake poop. Confabulation is a great word for it, and has a similar quality. But I suspect we'll end up having grown past the issue long before we'd see the term switch given the entrenched usage by now.
- B1FF_PSUVM 3y agoThanks, the description as confabulation is helpful, and should be more widely known. Maybe the current usage of hallucination can be viewed as metonymy or perhaps synecdoche - the untrained observer sees absurd outputs, the description of the phenomenon does not include the reason.
- sorokod 3y agoI like confabulation too but ultimately one way or the other we just anthropomorphizing the software. Does anyone seriously considers LLMs as modeling the human brain?
- dr_dshiv 3y agoWith confabulation it isn’t anthropomophizing but rather drawing a parallel to an information processing approach in brains. I don’t like hallucination because it implies that it is different from how it normally works. Theories of confabulation suggest that this normal brain behavior. Still not the best word.
- johnisgood 3y agoTo note, hallucinations (actually, they cause delirium) from anticholinergics are very different from hallucinations from psychedelics, which are closer to visual distortions instead.
- 65a 3y agoIt's arguable that those are not hallucinations, as there is not normally a misunderstanding about their (physical world) reality, just some amount of acknowledged distortion of perception of the real world.
- johnisgood 3y agoYeah, you are right, that is why I said that they are close to visual distortions, rather than hallucinations. To top it up, delirium-induced hallucinations differ from disassociatives (such as ketamine or DXM) as well.