Comment by impossiblefork

I agree that there's both cynicism and denial, but when I've explained my views I have usually been able to get through to the complainers.

Usually my go-to example for LLMs doing more than mass memorization is Charton's and Lample's LLM trained on function expressions and their derivatives and which is able to go from the derivatives to the original functions and thus perform integration, but at the same time I know that LLMs are essentially completely crazy with no understanding of reality-- just ask them to write some fiction and you'll have the model outputting discussions where characters who have never met before are addressing each other by name, or getting other similarly basic things wrong, and when something genuinely is not in the model you will end up in hallucination land. So the people saying that the models are bad are not completely crazy.

With the wrong codebase I wouldn't be surprised if you need a finetune.