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.
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.
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 (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.
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.
deleted by creator
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.
deleted by creator
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 (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.
deleted by creator