"Only" kind of misses the benefit though. I'm very bearish on "AI", but this is an absolutely perfect use case for LLMs. The issue is that if you describe a problem in natural language on any search engine, your results are going to be garbage unless you randomly luckboxed into somebody asking, with near identical verbiage, the question on some Q&A site.

That is because search is still mostly stuck in ~2003. But now ask the exact same thing of an LLM and it will generally be able to provide useful links. There's just so much information out there, but search engines just suck because they lack any sort of meaningful natural language parsing. LLMs provide that.

▲

another-dave 2 days ago | parent | next [-]

Speaking of which, could we apply vector embeddings to search engines (where crawled pages get indexed by their vector embeddings rather than raw text) and use that for better fuzzy search results even without an LLM in the mix?

(Might be a naïve question, I'm at the edge of my understanding)

▲

com2kid 2 days ago | parent | next [-]

> Speaking of which, could we apply vector embeddings to search engines (where crawled pages get indexed by their vector embeddings rather than raw text) and use that for better fuzzy search results even without an LLM in the mix?

Yes, this is how all the new dev documentation sites work now days, with their much improved searches. :-D

	▲	another-dave 2 days ago \| parent [-]
		ah cool right! I didn't know that. One for me to check out and understand more. Thanks!

▲

esafak 2 days ago | parent | prev [-]

Why stop there? The LLM can synthesize the results and spare you the work.

▲

another-dave 2 days ago | parent [-]

I'm talking about the scenario the GP referenced — where if you search for say "holiday" but get no results because the pages only use the word "vacation" which AFAIK is still a problem in regular search.

LLMs inherently would introduce the possibility of hallucinations, but just using the vectors to match documents wouldn't, right?

	▲	esafak 2 days ago \| parent [-]
		No, llms still use similarity search for candidate generation, unless you don't give them any tools.

▲

pluc 2 days ago | parent | prev [-]

"Instead of making search smarter we just decided to make everyone stupider"

Why invest in making users more savvy when you can dumb down everything to 5 year old level eh