3 ms·

It is increasingly incredibly important here to make a distinction between using an LLM with and without search. Without search, I agree with you. But e.g. Cha

It is increasingly incredibly important here to make a distinction between using an LLM with and without search. Without search, I agree with you. But e.g. ChatGPT with search enabled is often an invaluable research tool, and dramatically speeds up finding relevant sources. It basically automates the spidering of references and links, and also handles the basic checks for semantic relevance quite well, and this task requires little real intelligence or thought. Only once you hit a highly specific and niche technical domain will it start to fail you here (since it will match on common-language semantics that do not often align with technical usage). For a lot of topics, I now have the reverse feeling that a person who has NOT used an LLM to facilitate search—on basic or even intermediate questions—is increasingly more of a concern.

lrvick 1y ago
When someone is able to demonstrate they are more efficient than me in my areas of research and expertise by utilizing an LLM, I will at least pay attention to their workflow and investigate ways to replicate it with FOSS tools I can host, improve, debug, and control for myself. Until then, LLMs are of no interest to me. I am pretty good at using grep on local data caches.
- D-Machine 1y ago
  Yup, reasonable enough. In areas where you are a true expert (i.e. where there are only handfuls at most of people like you, or where all expert knowledge is not online and accessible by LLMs), you essentially already know where to look, and how to most effectively phrase searches, so a generic LLM that automatically crawls common databases and sites is going to be useless, almost by definition / necessity. Training is just fitting a manifold to training data (interpolating and extrapolating), but truly expert domains (as defined earlier) don't generally have enough data for these things to learn helpful semantics. Perhaps a model fine-tuned or specifically built for the niche domains might help non-experts and people learning to better do semantic / common-language searches, and this might accelerate learning. But, it might not. Not all domains have their core knowledge encoded in documents and/or language, and summarization is sometimes only harmful (especially for highly technical writing). It also seems to me almost definitional though that LLMs aren't going to beat true expertise, because LLMs aren't true intelligence or GAI, they really do just estimate token path densities / curve-fit. And real understanding is clearly more than that.

lrvick 1y ago

When someone is able to demonstrate they are more efficient than me in my areas of research and expertise by utilizing an LLM, I will at least pay attention to their workflow and investigate ways to replicate it with FOSS tools I can host, improve, debug, and control for myself. Until then, LLMs are of no interest to me. I am pretty good at using grep on local data caches.

D-Machine 1y ago
Yup, reasonable enough. In areas where you are a true expert (i.e. where there are only handfuls at most of people like you, or where all expert knowledge is not online and accessible by LLMs), you essentially already know where to look, and how to most effectively phrase searches, so a generic LLM that automatically crawls common databases and sites is going to be useless, almost by definition / necessity. Training is just fitting a manifold to training data (interpolating and extrapolating), but truly expert domains (as defined earlier) don't generally have enough data for these things to learn helpful semantics. Perhaps a model fine-tuned or specifically built for the niche domains might help non-experts and people learning to better do semantic / common-language searches, and this might accelerate learning. But, it might not. Not all domains have their core knowledge encoded in documents and/or language, and summarization is sometimes only harmful (especially for highly technical writing). It also seems to me almost definitional though that LLMs aren't going to beat true expertise, because LLMs aren't true intelligence or GAI, they really do just estimate token path densities / curve-fit. And real understanding is clearly more than that.