4 ms·
Again and again we see that LLMs are great for creative output and terrible for anything where correctness matters. You should only use it for the latter scenar
by bilalq 2y ago
Again and again we see that LLMs are great for creative output and terrible for anything where correctness matters. You should only use it for the latter scenarios when generating answers is slow/hard/expensive, but verification of answers is quick/easy/cheap. Probabilistic and non-deterministic answers have their place, but these companies marketing them in products need to do a better job expressing the limitations.
- mrweasel 2y agoIt shows an amazing lack of understanding for what an LLM is, even from the people selling and implementing them. You're exactly right in that they are terrible if correctness matters, but that should be obvious. If they where 100% correct, the size of the models would be much larger, as they'd need to retain all the original training data. You can use the LLMs for language understand and interpreting questions, but the would need access to databases containing authoritative answers and not answer anything for which they don't have an answer.