Hacker News Reader

Perplexity is definitely not sans hallucination and its accuracy varies highly depending on the type of query. For me it occupied a niche hybrid spot between full blown Llms and 'dumb' search engines but that niche is increasingly being squeezed out on both sides as the big Llms add web capability with more structured results while Google etc get smarter. It is hard to see what unique value proposition they offer going forward and its no surprise theyre becoming more google-like everyday, in the negative sense, as they struggle to justify their sky high valuation to all the vc money that is propping them up.

Well just depends how hard the questions you ask it are. If I'm getting help with graduate physics it can certainly combine near-miss sources into a mishmash hallucination. Granted, the sources can be useful but they're pretty exclusively things I've already found via google. I think only one has it returned a novel, useful source

I never ask graduate level Math, Physics, or AI questions to any LLMs. I don't trust them enough.

This depends entirely which model you choose right? Which model are you using? I alternated between Sonar Huge and Claude 3.5 Sonnet when I used Perplexity (sometimes switching on a per-query basis), but the decision fatigue with model selection really got to me.