@louis Nice! Thanks for a nice showcase of another obvious limitation of LLMs.
Hallucinating is bad enough, but to completely fabricate new explanations on non-existing stuffs really take the cake here (but not too surprising; since fundamentally there is no concept of "truth" in LLMs anyway).
This reminded me of another case where a user asked an LLM a technical question, and it give a very plausible answer; only for the user to realized later that the correct answer is the complete "reverse" of the one that the LLM gives.
This is especially hilarious because I know a couple of companies that is building their product around Mistral AI.
Can't wait for the bubble to burst
EDIT: And no, RAG won't save LLMs; people already have to downplay the hype of what RAG can do to mitigate issues with LLMs.
@olav And this is what ChatGPT has to tell about HUJU Common Lisp on my machine.
Here you can see that LLMs actually hallucinate depending and what it predicts you want to know.
Apparently "Does it exist?" is different from "Tell me about property N of X". Because in the latter case it assumes that it already exists.
To be more precise, when I refer to AI in the Mastodon discourse, I mean LLMs. There are legitimate use cases for Machine Learning.
The problem is how LLMs are sold to the world as actually useful while they are utterly broken. And a society that relies on broken software ultimately becomes a broken society.
@louis I don't really know. That discussion quickly becomes a philosophical discussion about what knowledge even is.
In practice I find ChatGPT4 very useful for questions in a broad range of areas. As opposed to reading articles, books, docs etc, it can tailor the info for you, so that if you ask the right question you get just the right info back instead of having to scavenge and piece together what you want to know. Especially useful if you don't know exactly what you're looking for.
fwiw my feeling from talking to it is that it has a lot of knowledge.
I wonder, do LLMs actually have knowledge of the real world, or is it just a giant text transformation engine?
Idk
using LLM along with knowledgebases is quite popular. I think it's called RAG?
Basic example would be passing a whole wikipedia article about the subject to the llm...
And a society that relies on broken software ultimately becomes a broken society.
I think of it as the inverse: a broken society produces broken software. Why would a society such as ours be so driven for this crap? Your point still stands however.
@olav I see know that you use ChatGPT 4, while I used ChatGPT 3.5. However in my former examples I was using Mistral AI's LLM - which is praised specifically in the Fediverse as being "superior" to ChatGPT.
Add comment