People are drawn to things that look cool and innovative for the sake of it, but we're not good at thinking ahead to how well they'd work in practice.
I've been searching the web for so long that it's like breathing. I type a few keywords, hop between the first few results very fast, get what I need and get out. Maybe I have to repeat. Few searches are more complicated than that.
I've tried replacing that flow with AI. It's slower to input, slower to get results, the results are opaque with regards to where they came from and how trustworthy they are. And the results aren't better, in my experience as someone for who finds search to have been solved and painless for many years.
Even asking one question of an LLM takes more time and effort than a traditional search, and the problem is compounded when I have to iterate on a search.
(Also it sucks to cede the kind of power to search providers that gives them leverage to force us to be logged in and monetize what's always been free, but those are just random complaints.)
So, going back to your question. My guess is that Google has always known this isn't a better overall experience, once you get over the initial wow factor. But maybe they underestimated how much that wow factor would allow true competition to materialize before the rest of us would figure it out and lose interest.
So far, the only use-case I've seen for LLMs are basically code completion.
It would need the ability to contain infinite context to be genuinely useful in my opinion.
There are dozens if not hundreds of other uses out there.
A search engine is a great peripheral for a language processor. Bing chat does show how you’re supposed to connect the two, as are ChatGPT plugins.
LLM can do synthesis and summaries of multiple search results. It is slow to get a response and the thing makes mistakes sometimes but we’re in the first quarter of the tech being deployed. It’ll only get better.
I've been using DuckDuckGo as my daily driver for years. It's generally OK. It sucks, but just barely works well enough to keep me going.
I started using Perplexity and ChatGPT for some searches that DDG was having trouble with, and it blew me away with the quality
Then I started going back and trying the same searches on Google, and I was surprised to find that it was just as good as the LLM-powered results after all. I don't know if phrasing things as questions was the key or just using Google instead of DDG did the trick, though
Call this a 90/10 split.
Of that 90, I would say about 75% of my questions can be answered by GPT* trivially. Most of this is due to the fact most of my problems are not novel.
Of that 10, the results are much more all over the place. I would say half or less.
The thing is, when you get crappy results from the LLM, they are genuine technical limitations (as long as you avoid politics and religion, if you are researching those with the LLM, good luck with that), not intentional badness put there to make someone else a few pennies off of wasting my time. There is a certain value there that is hard to quantify.
The other killer feature is being able to have a conversation about the results. This, for me, pushes the usefulness factor way above Google. The options for having conversations about technical topics with humans are comparatively god-awful.
You're probably not using it right. :)
Its utility as "search" is only partially overlapping with search. It's most useful for that when you can describe what you want but not with clear enough terms to search for it. For research now, I'll try the search engines-- if I don't get hits I'll ask GPT4. Even if it gives me a wrong answer it'll almost always use terms that let me form a more useful search, though often it just gives right answers.
Simple example, last night a old friend emailed me to ask if I remember the name of a 90s website that people used to brag about linux uptime and deployment. He couldn't find it via search. I remembered what he was talking about, couldn't remember the name, also couldn't find it via search. Asked gpt4 (by simply pasting in the email), it answered that it was probably the "linux counter".
It's possible that with more guessing and grinding on search terms I might have hit the Wikipedia article on it, but GPT4 gave an immediate and accurate answer and included the URL which is what I needed to load the page up on the internet archive (as the site is long gone from the internet).
The main thing GPT4 can't do is any kind of exact string search -- but unfortunately google more or less won't do that either anymore. In terms of figuring out what you intend from some vague text GPT4 usually does vastly better than google.
I think my ideal tool would in parallel give me an not-very-interpreted classical search result and a state of the art LLM result. It seems that Google's bard works kind of in that direction, but their LLM seemed extremely dumb and hobbled when I tried it even compared to GPT3.5 and the search elements are google. ... so worst of all worlds. :)
PS. Just read the Lemoine's transcriptions. Just wow.