Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

ChatGPT won’t replace search engines any time soon

algolia.com

Post reply on HN
Print view View original

141–150 of 427 posts

Re: ChatGPT won’t replace search engines any time soon

#141

Post by ben_w »

schrodinger wrote: post #138
ben_w wrote: post #65

Earlier quoted context omitted.

I've seen this sort of response a lot lately, both for GPT and all the image generators. Google, and all the other search engines, literally show snippets of the original content to display search results, and have had copyright infringement lawsuits over this. Yet somehow, it's taken these AI, which generally aren't just block-quoting things unless you ask them to (and often getting it wrong if you do), for non-lawy…

Is it any different than an infinitely capable human reading the entire internet and using the knowledge it gained to answer questions?

Google or GPT?

Either way I don't know. Might be the same, but that might be an unwarranted anthropomorphisation.

Re: ChatGPT won’t replace search engines any time soon

#142

Post by ranting-moth »

iLoveOncall wrote: post #112
ranting-moth wrote: post #93

They might not replace them, but it fulfills a role that Google could do previously. Try to find out what type of a spring is in a spring clamp using google? "What type of a spring is in a spring clamp?" Or use whatever query you want. The top 10 results won't tell me the answer. Ask ChatGPT. It'll tell you (spoiler: it's a tension spring or a torsion spring. It's not a coil spring.) 8 out of 10 Google results are tr…

> Ask ChatGPT. It'll tell you Without any way for you to know if it's the right answer or not. Wow, so useful.

Well, I then did a google image search (which is much better than it's traditional search) and it came up with things that did look like that one I had in my hand.

But I do get where you're coming from, ChatGPT is way too confident when it's wrong!

Re: ChatGPT won’t replace search engines any time soon

#143

Post by dustincoates »

tomlue wrote: post #11

Article claims that chatgpt is overly sensitive to queries eg -gate suffix and come lad vs cum laude. This seems like a weak argument based on a temporary shortcoming. Any NLP methods search engines use for word meanings should be trivially integrated into chatgpt. Comes off as a company that sells search solutions looking for reasons this tech won't replace it. I'm sure the author is better informed, but what about…

Author here. I was surprised to see my post on here while scrolling!

My point about those examples is less that ChatGPT got it wrong, but that it's impossible to know really when it does, because it seems so confident and you only get the one result. To be fair, this is a problem with webpages, but you at least get multiple choices.

> Comes off as a company that sells search solutions looking for reasons this tech won't replace it.

I can see why you'd say that, but that wasn't the goal. I originally wrote this as my own POV for my own blog. I really do think there's a UX problem in this (and, as I mentioned, I am a big fan of recent LLMs).

> [What] about the scalability of search?

You're right, the NN aren't as performant as lexical search. They're getting a lot better though. (We're actually working on this at the moment.) LLMs, though, have a ways to go, so it's hard to use them right now for real time search.

Re: ChatGPT won’t replace search engines any time soon

#144

Post by ranting-moth »

wizofaus wrote: post #121
ranting-moth wrote: post #93

They might not replace them, but it fulfills a role that Google could do previously. Try to find out what type of a spring is in a spring clamp using google? "What type of a spring is in a spring clamp?" Or use whatever query you want. The top 10 results won't tell me the answer. Ask ChatGPT. It'll tell you (spoiler: it's a tension spring or a torsion spring. It's not a coil spring.) 8 out of 10 Google results are tr…

Curious, the second result for me was aboutmechanics.com, which said it was a coil clamp. But I agree Google seems to be interpreting the question as "what's a spring clamp".

Interesting. aboutmechanics.com is not in my top 10 results. But even aboutmechanics is wrong, it's not coil spring. It's torsion spring.

Re: ChatGPT won’t replace search engines any time soon

#145

Post by dustincoates »

varunkmohan wrote: post #99

This feels like what a company that hasn't accepted the future would say. Yes, ChatGPT doesn't do exactly what Google does. Can it be augmented by well understood search retrieval engines to generate a much better response? I think for a lot of cases that seems to be undoubtedly true.

I'm (the author) actually in agreement with you. LLMs are going to be a big part of search in the future. I alluded to that I'm the post. I'm less convinced about search as a chat interface. But LLMs for query understanding, ranking, etc.? Of course.

Re: ChatGPT won’t replace search engines any time soon

#146

Post by iLoveOncall »

sgc wrote: post #134
qayxc wrote: post #126

Earlier quoted context omitted.

Google results(plural!) give you alternatives and references, i.e. you know where each answer is coming from and SEO spam is easily identifiable. ChatGPT just gives you an answer that's confidentially wrong sometimes.

But isn't that just a UI problem at heart? It would be easy to ask for references for its conclusion.

ChatGPT doesn't have references and doesn't make conclusions, it's a statistical model.

Ask it for references, it will invent you URLs that look legit but don't exist.

Re: ChatGPT won’t replace search engines any time soon

#147

Post by raegis »

skybrian wrote: post #35
raegis wrote: post #16

Just heard of this, so I logged in with my google account. The mistake it makes below is interesting. If this were a human response, I would assume a brain fart. Not expecting this type of error from "A.I." Me> Find the area of an ellipse with semimajor axis 5 and semiminor axis 4. ChatGPT> The area of an ellipse with semimajor axis 5 and semiminor axis 4 is 40π. You can find the area of an ellipse using the formula:…

The trick is to get it to show its work first and then calculate the answer. If it guesses the wrong answer first and then writes out the steps, it won't correct it. This works: "Find the area of an ellipse with semimajor axis 5 and semiminor axis 4. Don't calculate the result until the end."

After computing your example, I tried my original question which now gives the correct answer. In fact, it gives essentially the same response for both forms of the question. I guess the training works as it should.

Re: ChatGPT won’t replace search engines any time soon

#148

Post by antihipocrat »

aflag wrote: post #107
antihipocrat wrote: post #80

Earlier quoted context omitted.

If we don't need to click a link does google receives less revenue? I think including more direct answers from a search query would require changes to Google's business model.

The ads would still be clickable. Clicking the actual results doesn't really help google much.

Except for any ads embedded in the referred sites.

Re: ChatGPT won’t replace search engines any time soon

#149

Post by notahacker »

pattyj wrote: post #135
notahacker wrote: post #87

Earlier quoted context omitted.

> And as a logical aside: if AI can't be trusted, we shouldn't trust AI's recommendations for reliable sources either. Don't think that's automatically true though: an AI can be consistently excellent at retrieving reliable sources and still get a lot wrong on its summaries. Also humans have their own opinions on sources even when they're not familiar with the detail, and finding out if the AI reads Nature or 4chan i…

Sure. Really, I was thinking of the scenario where someone explictly asks for a reliable source i.e. "Thanks for summary about X. To whom should I turn for more accurate/detailed information?"

Oh yeah, it's spectacularly bad at some of that sort of thing. Ask for a list of academic papers and it'll make half of them up and give others new authors...

Re: ChatGPT won’t replace search engines any time soon

#150

Post by raegis »

tarl0s wrote: post #40
raegis wrote: post #16

Just heard of this, so I logged in with my google account. The mistake it makes below is interesting. If this were a human response, I would assume a brain fart. Not expecting this type of error from "A.I." Me> Find the area of an ellipse with semimajor axis 5 and semiminor axis 4. ChatGPT> The area of an ellipse with semimajor axis 5 and semiminor axis 4 is 40π. You can find the area of an ellipse using the formula:…

Being not very good with numbers is one of the limitations of the current model. I imagine that in the future we will see integrations with a number of different tools to overcome these limits - WolframAlpha being the first one that comes to mind when talking about algebra and calculus.

A friend told me students were using this for cheating on math exams, so I assumed (incorrectly) it had something like Wolfram Alpha built in.
Post reply on HN