It's another tool. Learn to use it, and learn when it shouldn't be used.
Google abandoned the ability to use most search qualifiers effectively, so for niche websites I'll get zero results even with perfect exact-string-match queries, even when the site is in Google's index. On the other hand if I only vaguely remember the content and no related words or synonyms, Google is unable to turn my fuzzy feelings into an accurate search result. Plus, your ability to filter the "kind" of site is basically non-existent, making it all but impossible to find information on topics vaguely related to any word vaguely related a product you could potentially buy.
LLMs hallucinate and have their own set of problems, but that orthogonality makes them very useful situationally.
Not too long ago I needed to track down the blog Google references internally in the design doc of their TGIF employee voting platform, regarding Wilson scoring (using confidence intervals instead of means/medians/...). That's very easy to do in Google search if you can remember the right keywords from a decade ago (like Wilson scoring), but otherwise it's impossible. Reframing that problem for an LLM, you'd use plain English to describe everything you know, add a bit of flattery to shift the output distribution to that corner of the internet which actually knows what you're asking, potentially add one sentence to stop this last batch of models from wasting their time actually searching the web, and ask for a list of the top 5 authors and blog titles they think might be correct. That gives you a whole new set of search terms to finish your quest (in my case, the right answer was always in that list of 5, no matter how many times I reran the query).
That property of having to add additional context (e.g., when asking for recipes, I'll describe the background of who the LLM is roleplaying first) to get a good result is annoying. Full sentences with proper punctuation, unfortunately, also help. I wrote a small tool to make it easier for me to keep track of prompts I found useful and execute them with modifications.
As some other commenters mention, the LLM can help you with XY problems. You're searching for recipes, techniques, nutrition spreadsheets, ..., trying to craft something that meets some set of constraints (e.g., for some hypothetical set of guests you might require: no pork, most dishes have to be vegetarian, most dishes have to be gluten-free, it's fine if cooks all day so long as that isn't active prep time, you'd prefer to make it as tasty as possible while leaning in to cheaper, homestyle cooking, it has to use these red bell peppers I have already, and nutrition doesn't really matter). That's a nontrivial collection of tasks with a proper search engine, almost all of which the LLM is piss-poor at individually, unless you already have a good idea of the kind of dish you want to make. However, if you in two phases ask the LLM to brainstorm a list of 20 meal ideas and then expand on your favorite (that would be a decent time to add any modifications) then that collection of tasks gets done all at once. You _have_ to be able to look at the recipe and decide if it's any good or not, so a beginner probably shouldn't do that, but for everyone else it's a huge time saver.
I mentioned that Google sucks at filtering the "kind" of website you'd like to visit. LLMs handle that great. Like always, you have to be able to handle hallucinations (in this case, by just going to the results and checking if they're any good), but consider a prompt like the following:
> The web nowadays has tons of ad-infested, profit-driven, barely legible SEO drivel -- even in the top 100 search results and even from huge sites -- but all the old websites like Sheldon Brown's wealth of bicycle knowledge are still out there. List the top three old-web resources I'd want to read to learn about grafting apple trees.
When I ran that, I got one commercial result, one journal with a wealth of paywalled information, and one forum with a huge collection of free information about the particular species of Apple I'm interested in.
For "apple grafting" in particular, Google does just fine (a bit better arguably since the 1st Google result was the 2nd LLM result, and it's the only one I really cared about), but the more commercialized the knowledge you're looking for is the more the LLM shines out in comparison.
Google abandoned the ability to use most search qualifiers effectively, so for niche websites I'll get zero results even with perfect exact-string-match queries, even when the site is in Google's index. On the other hand if I only vaguely remember the content and no related words or synonyms, Google is unable to turn my fuzzy feelings into an accurate search result. Plus, your ability to filter the "kind" of site is basically non-existent, making it all but impossible to find information on topics vaguely related to any word vaguely related a product you could potentially buy.
LLMs hallucinate and have their own set of problems, but that orthogonality makes them very useful situationally.
Not too long ago I needed to track down the blog Google references internally in the design doc of their TGIF employee voting platform, regarding Wilson scoring (using confidence intervals instead of means/medians/...). That's very easy to do in Google search if you can remember the right keywords from a decade ago (like Wilson scoring), but otherwise it's impossible. Reframing that problem for an LLM, you'd use plain English to describe everything you know, add a bit of flattery to shift the output distribution to that corner of the internet which actually knows what you're asking, potentially add one sentence to stop this last batch of models from wasting their time actually searching the web, and ask for a list of the top 5 authors and blog titles they think might be correct. That gives you a whole new set of search terms to finish your quest (in my case, the right answer was always in that list of 5, no matter how many times I reran the query).
That property of having to add additional context (e.g., when asking for recipes, I'll describe the background of who the LLM is roleplaying first) to get a good result is annoying. Full sentences with proper punctuation, unfortunately, also help. I wrote a small tool to make it easier for me to keep track of prompts I found useful and execute them with modifications.
As some other commenters mention, the LLM can help you with XY problems. You're searching for recipes, techniques, nutrition spreadsheets, ..., trying to craft something that meets some set of constraints (e.g., for some hypothetical set of guests you might require: no pork, most dishes have to be vegetarian, most dishes have to be gluten-free, it's fine if cooks all day so long as that isn't active prep time, you'd prefer to make it as tasty as possible while leaning in to cheaper, homestyle cooking, it has to use these red bell peppers I have already, and nutrition doesn't really matter). That's a nontrivial collection of tasks with a proper search engine, almost all of which the LLM is piss-poor at individually, unless you already have a good idea of the kind of dish you want to make. However, if you in two phases ask the LLM to brainstorm a list of 20 meal ideas and then expand on your favorite (that would be a decent time to add any modifications) then that collection of tasks gets done all at once. You _have_ to be able to look at the recipe and decide if it's any good or not, so a beginner probably shouldn't do that, but for everyone else it's a huge time saver.
I mentioned that Google sucks at filtering the "kind" of website you'd like to visit. LLMs handle that great. Like always, you have to be able to handle hallucinations (in this case, by just going to the results and checking if they're any good), but consider a prompt like the following:
> The web nowadays has tons of ad-infested, profit-driven, barely legible SEO drivel -- even in the top 100 search results and even from huge sites -- but all the old websites like Sheldon Brown's wealth of bicycle knowledge are still out there. List the top three old-web resources I'd want to read to learn about grafting apple trees.
When I ran that, I got one commercial result, one journal with a wealth of paywalled information, and one forum with a huge collection of free information about the particular species of Apple I'm interested in.
For "apple grafting" in particular, Google does just fine (a bit better arguably since the 1st Google result was the 2nd LLM result, and it's the only one I really cared about), but the more commercialized the knowledge you're looking for is the more the LLM shines out in comparison.