• This is a useful take: I too will use LLMs for search-- but not for search for journal articles with data and evidence. LLMs too easily confabulate these.

    LLM-as-search is fantastic when you want a no-bullshit statistical result for what you’re looking for when you’re wanting an overview or interactive tutorial.

    • As long as it has footnoting so I can see where each piece of information was sourced from, AI chat has its use cases. Without that I genuinely do not see the point at all. It’s like when people “ask Google” something and just blindly trust the highlighted “answer” as infallible truth. It’s just a really, really bad habit to develop and I wish more people understood this.

      •  Zworf   ( @Zworf@beehaw.org ) 
        link
        fedilink
        1
        edit-2
        7 months ago

        Not infallible truth. But very often it’s something that is just for personal use.

        Some things I’ve asked it recently were like “Which torch is smaller out of these 5 models?”. Once I find which one I want it’s easy to verify. Or “what does this Spanish expression mean?” or “how do I do …”.

        Not everyone uses it to try and write authoritative stuff. And Google is full of clickbaity “comparison sites” that are nothing but fake advertising.

          •  Zworf   ( @Zworf@beehaw.org ) 
            link
            fedilink
            1
            edit-2
            7 months ago

            Yeah but accuracy isn’t a given with the other methods either. If I ask some randos on reddit I won’t get a perfect answer either. If I google specs or reviews online they are often biased, wrong (think the magical Chinese lumens of torches) or even literally fraudulent paid reviews too.

            So yeah for me the LLM output is more than good enough with a bit of verification if necessary.

            I don’t really understand why people are suddenly hung up about holding LLMs up to this lofty ideal of an unbiased super-truth. Where did that requirement come from all of a sudden? It’s not really realistic and not something we’ve ever had in the past.

            I feel the same about self-driving systems. People get all hung up if they crash once in a while, expecting them to be 100% perfect in all situations. But ignoring the concept that they already might be a hell of a lot safer than human drivers. They fail in different situations generally but why do we suddenly demand perfection?