4 ms·
With LLMs hallucinating and being generally unreliable, I'm coming to the conclusion that LLMs should only be used to transform natural language to structured d
by kgeist 3y ago
With LLMs hallucinating and being generally unreliable, I'm coming to the conclusion that LLMs should only be used to transform natural language to structured data (specific to your application), and knowledge about the world should be stored somewhere else (some sort of vector databases? tools?), and, fortunately, smaller models are already quite good at the former! Trying to extract world knowledge from LLMs can be a deadend... I don't see how the hallucination problem can be 100% fixed for LLMs themselves.
- tjr 3y agoI do not trust LLMs. They may very often be right, and may very often be useful, but I do not trust them. If you require correct results, their output should be checked somehow.
- aeternum 3y agoStrange take, do you consider world knowledge produced by humans to be anywhere near 100% accurate? Suppose you had an oracle that when asked a question gives you 5 answers, 1 of which is true. Or even 1 of which is true only 50% of the time. You would still generally be a fool to throw away that oracle.