5 ms·
> In Google's case they are still really focused on search whereas LLMs arguably move the focus to answers. I would love to see what proportion of searches are
by kweingar 3y ago
> In Google's case they are still really focused on search whereas LLMs arguably move the focus to answers.
I would love to see what proportion of searches are questions that would benefit from natural language answers. The huge majority of my searches would not be improved by LLMs and in fact would probably be made worse. “Thai food near me”, “IRS phone number”, “golang cmp documentation”
- throwaway1777 3y agoThai food near me- chatgpt can give you a list of thai restaurants near you. IRS phone number has a definitive answer. Chatgpt can also spit out the golang documentation for cmp or even give you sample code.
- kweingar 3y agoOk but what exactly is the benefit of using ChatGPT for that? It is more than a year and a half out of date. It doesn’t know the Thai place’s current hours and doesn’t automatically surface a one-click link to their menu or reviews. Why would I use ChatGPT to get the IRS phone number when I could use just as little effort typing it into a search engine and going to their actual .gov site with no risk of hallucinations? When I’m using a new library, often I want an overview of the types and functions it surfaces. Why would I use an outdated and possibly halluciniated answer (that takes 30 seconds or more to generate) instead of clicking on the link in Google and having a nice document full of internal links appear instantly? I don’t want to use the chatbot for the sake of using the chatbot. I want to use it when it’s better than what I already have. Sometimes it is better, and in those cases I use it a lot!
- known 3y ago[dead]
- re5i5tor 3y agoSolved problem. The solution is NOT ChatGPT, it’s something like Bing or Phind that is a hybrid of traditional search and LLM.
- kweingar 3y agoHere is my experience using ChatGPT to find local Indian restaurants. The responses took about 90 seconds in total to generate and gave me very little information. Why would anybody use ChatGPT instead of a search engine for this kind of thing? https://ibb.co/pdzdncZ https://ibb.co/pdzdncZ https://ibb.co/LRP6McY https://ibb.co/LRP6McY Compare to Google/Bing which took about 5 seconds to type in the query and return these results. https://ibb.co/1R48Bwc https://ibb.co/1R48Bwc https://ibb.co/P4XGXxW https://ibb.co/P4XGXxW
- eastern 3y agoYeah for questions like this ChatGPT sounds like someone who's being paid by the word. Always says 200 words where 30 would suffice.
- ryanmerket 3y agoBecause the average American is considered to have a readability level equivalent to a 7th/8th grader (12 to 14 years old). They lack the critical thinking skills to from from search results to prioritized list. :-/
- bradknowles 3y agoFor decades, the standard was 6th grade. I believe that was part of the APA guidelines. But at least Americans have been getting stupider, on average. IMO, now I think you need to write at the 4th or 5th grade level.
- colberding 3y agoReframed as "Why would anybody use ChatGPT instead of a search engine for this kind of thing" today? You're correct - ChatGPT is missing some bells and whistles such as real-time data, your location, how many times you've visited the websites of certain restaurants, and so on. However, so do 'search engines' in their base implementation of indexing and ordering links to other websites. I think you'll see some technical limitations (real-time data) overcome and user-focused implementations and features of AI/LLMs emerge over the next year. At that point, I think your initial question becomes more relevant in a general way.
- cableshaft 3y agoI asked it to generate a travel plan including restaurants for a trip I’m considering. What it generated included some places that were now closed, but it was an excellent starting point, and it beat my usual approach of a Google search of the area and tapping random places in the area.
- intended 3y agoFrankly, it’s more about the number of ads and low relevance. Old google was simply faster to use. GPT for search is google without ads
- kweingar 3y agoThe UX for ChatGPT is awful compared to search engines, at least for quick, easily found facts like what I mentioned.
- ryanmerket 3y agoDepends on the fact. If you ask Google for the difference between Viet red tea and Viet green tea, ChatGPT can give you the correct facts much more quickly than Google.
- BlueTemplar 3y agoDepending on the facts, you can just directly search Wikipedia (or other more specialized websites) - I have had it on the w keyword for nearly two decades now...
- sdwr 3y agoThat kind of myopic thinking is exactly why google might be in trouble. Think about the problem, not your current solution. "I'm hungry" "I need to do my taxes" "My code dont work right" Searching for info is a solution to those problems, not the solution. The promise of AI (might take a while to get there) is having an agent that you trust to solve those problems for you. Or learning your preferences over time. Or folding the question into part of a longer dialogue.
- thaumasiotes 3y ago>> The huge majority of my searches would not be improved by LLMs and in fact would probably be made worse. “Thai food near me” > Think about the problem, not your current solution. > "I'm hungry" This only convinces me that you didn't do any thinking about the problem.
- mk89 3y agoI agree with this comment. We have become great googlers: how to search for things to solve problems. It's not different from a mega huge yellow pages well structured etc. Next is: how to ask for help to solve some problem.
- splistud 3y ago[dead]
- derefr 3y agoBut when googling, I, the human, am often already acting as an "agent that you trust to solve those problems for you" for some higher-level question-asker. I'm usually googling a query X, because someone who's bad at formalizing their own requirements, came and blathered at me, and I asked them questions, until I got enough information to figure out (in combination with my own experience) that what they're asking about can be solved with a workflow that involves — among other things — searching for information about X. (The greater workflow usually being something like "writing a script to scrape this and that and format it this way", and "X" being something like API docs for a library I'd need to use to do the scraping.) Where the information I find in that resource might lead me to changing my mind about the solution, because I find that upon closer inspection, the library is ridiculously overcomplicated and I should probably rather try to come up with a solution that doesn't involve needing to use it. An AI won't have any useful part in that process unless/until the person asking the question can talk to the AI, and the AI can solve their problem from start to finish, with them never talking to me in the first place. Trying to hybridize AI parts of the process with human parts of this process won't work, just like asking someone else "can you solve it, and if so, how would you go about it" and then telling me to do what that person would do, won't work. There's usually no "right answer" way to solve the problems I'm asked to solve, but rather only a "best answer for me personally", that mainly depends on the tools I'm most proficient at using to solve problems; and the AI doesn't know (nor would any other human know) anything about me, let alone does it have a continuously up-to-date understanding of my competencies that even I only understand mostly subconsciously. So it can't apply my (evolving!) proficiencies in these skills as constraints when deciding how (or if!) a given problem can be solved by me.
- agitator 3y agoAnd isn’t that the problem? Why do I need to translate my question into an optimal set of keywords that will give me what I want while minimizing unwanted results? Google search was a great stepping stone and connects you with the web, but it’s broken in many ways when it comes to what value we are really trying to extract. A machine that can hone in on what I’m getting at in an intuitive sense while having all of human data available to generate a response is so much more powerful.