No matter what, Perplexity is now the worst of the worst. They were early with the ability to upload documents but the utter failure of Perplexity to be useful is proving what I have been saying for a year now, (1) LLMs are not "AI" any more than a spell checker is and (2) LLMs are not really useful for anything other than a VERY good spell checker.
> not really useful for anything other than a VERY good spell checker.
Try GitHub code pilot. Try asking it to summarize something you know nothing about and then tell it that you need the explanation to be more simple several times. How can you compare this to a spell checker?
People keep trying to apply them as autocompleters... that doesn't mean they're good at it.
As to why an LLM is not AI? People are wrongfully "afraid" that an LLM is somehow make "bad" information too easily available to people. Like how to rob a bank or make a bomb. The confusion is in thinking that an LLM can output something that it had not seen in training and supposedly all the training data is available elsewhere. Now, "data leaking" is a serious issue with LLMs but that has nothing to do with "AI".
An LLM is a good auto-complete, which is a good spell checker, which is a good search algorithm. I don't see how we ever get from context search and retrieval to "AI".
Do you think (for example) someone asking you what you are would give them enough ability to judge your abilities? Yes you would get the answer right and the models you are choosing are getting the answer wrong, but you can't extrapolate from that to anything else at all about what this model can or can't do.
It’s a little silly to test a small LLM with a question that at least requires knowledge of its own construction that was not included in its training set, and which really requires existential introspection.