I've replaced search engines with Wikipedia as the first step in looking for information. skimming an article is much faster than opening and closing a dozen sloppy s.e.o. 'd websites :p
AI overview uses around 20 times more energy than regular searches.
It's absolutely batshit crazy that most search engine put it on by default just for it to summarize a wiki page 90% of the time.
It's like we're collectively pushing for a mass extinction.
What's even crazier is for a lot of things we already had a quick result overview. Like if you wanted the definition of a word. For a long time it was the case that you could just "define socks" and you'd get a widget at the top with the definition. Why would you replace that with AI, and have an AI process that request every single time someone wants to define socks instead of just keeping the old widget!?
I'm pretty sure a good percentage of the searches are also just people typing the name of the website they want to access in the search bar. They don't want any information, just a link. All these searches now consume a massive amout of energy for basically nothing.
If it didn't do the AInsearch repeatedly they couldn't use the billions of daily searches to artificially boost their AI engagement metrics and further inflate this bubble.
Me after losing 6 days of testing broken theories from an ai search when I could have just spent an hour watching a video that a human made that was fully functional from the start.
That said: No doubt that human spent way longer discovering the technique and then making the video: this should also be considered as a reality that creating something good takes time. And humans are still better at creating scalable content just give them the proper R&D time ffs.
LLMs fundamentally can't learn. They will continue to fail the same thing over and over until they guess and check into a response that appears coherent and correct.
All jokes aside, I usually do want to read the six Wikipedia pages because getting a one-sentence answer with no context is risky and not all that useful anyway. I need the nuance and background info for any question worth asking.
If you need nuance and background, wikipedia is NOT the place to get that. Maybe you can get some good links at the bottom of the page to real information, but that can be hit or miss also.
I would say that even simple answers at least deserve a source or citation. This is my biggest complaint with AI search, you get an "answer" with no qualification. The citations listed in the AI summary often don't contain the actual info being presented.
AI has never been 100% accurate to the documents it pretends to reference, so the fact that anybody could even consider reading it like it's not pure gibberish is mind boggling to me.
I gave Claude a document to proofread today. It came back with an error. Clause said I used "told" instead of "tell." The error ended up being Claude's error, as I used "tell," and it was just confused.
This was the first time I ever used Claude. Just wanted to see what it was like. It did apologize, in fairness. But yeah, it's clearly fallible. I guess I don't understand it well enough to get how it read the document incorrectly. I'm baffled.
It's because it's a statistical model that guesses the next word. You asked it to read a document and formulate a response. So, based on responses in the dataset to the documents in the dataset, it chose the words and phrases which were most likely to be in a response.
Statistically, the most likely result the machine could come up with, was that you had used the wrong word. Factually, you did not.
Lmao, I guess my internal ad filter is still working because I had to look at that twice to notice it and wondered if there was a new Halloween movie coming out this year or something. Agreed, fuck that shit.
Yesterday there was a weird play in the Thursday Night Football game where the Bears' quarterback handed the ball to an offensive lineman, who ran it in for a touchdown. The referees threw a flag and called the play back because it was an "illegal handoff", which I had never seen before.
I was texting my brother during the game and we both commented on how weird it was. My brother googled it and read the AI summary, which said the play was legal - citing a reddit thread from several years ago. He was like "Google says the play was legal - the refs got it wrong!"
I had to go to the NFL rulebook and dig through a bunch of sections before confirming that the refs were correct and you can't do that.
AI is absolute dogshit for anything that requires a bit of nuance. But it is really fast at serving up that dogshit, so a lot of lazy people would rather just accept a quick wrong answer than a slow correct one.
It would have been a legal play IF the lineman had checked in with the referee as an "eligible" receiver before the play. But evidently he screwed up and report in. It's not a common play in the NFL, but every team has such a play on hand.
Yeah exactly. The AI was just wrong and said it's legal as long as you hand it off behind the line of scrimmage. I actually still have the screenshot he sent. Notice the 2nd bullet point.
It's not just wrong. It's plausibly wrong, which is worse. It's the difference between giving someone a recipe for cookies where you fuck up the ratio of flour and sugar vs giving someone a recipe for cooking where you've jokingly put "50 pounds of gunpowder" in there as an ingredient. People know the obvious wrong thing is wrong. But they'll bake the fucked up ratio cookies and then just not know what went wrong.
They also give you the option to (do their best at) blocking AI-generated search results. (Not perfect, of course, but it does a decent job of reducing how much AI slop you get in your search results.)
The funniest part is I've gotten better answers from duckeuckgo and Bing's AI back in the day because they both didn't treat reddit as a reputable source of info lmao.
Dunno how the scene is now, but I wouldn't be surprised if its the same.
THIS! I recall somebody doing contract review and having AI summarize the contracts just to see how well they did... The AI left out some really important stuff and also occasionally hallucinated stuff that wasn't there. Summarizing a contract should be pretty black-n-white -- if AI can't handle that properly, I'm not going to trust its summary of something more complex.
That's right. I didn't put the question in the search bar for you to tell me the answer. That question was the first shoveling of dirt in a 6 hour rabbit hole that I was fixing to explore.
Yep. I still have yet to see an actual use for AI. People are like “but summaries!” Summarize this: summaries are only useful if the information being summarized doesn’t matter, which to me means you could just skip the entire thing.
This isn’t a high school test, this is real life. Why are you trying to use SparkNotes?
Fake product reviews, for example. AI is fantastic for that. Want to buff up your shitty product's Amazon reviews? AI can generate the text for those reviews perfectly well.
As someone with ADHD, I disagree. Sometimes the information is important, and I don't want to have to dig through a wall of text to find it. One thing that annoys me about normies when they write, is that they tend to blab on forever on things that don't matter instead of getting straight to the point.
If people wouldn't use so many unnecessary words, I wouldn't need a god damn LLM to summarize their ramblings into something my brain can actually parse.
They are great to tell you the words you need to search for something; they are great for analyzing "vibes" out of texts; they are reasonably good to generate you a link to some site that answers a hard to find question, traditional search would be better if there were any one around that isn't completely broken.
I know the rankings change like every month, but isn't Gemini one of the "dumber" ones? I've only tried it a few times and got annoyed, but it's been a while.
That's still full of AI, the filter quality you can see for yourself without an account for a limited amount of searches. If I search an image and the URLs on slop I see in the results with the slop filter enabled that literally contains keywords related to AI slop and KAGI doesn't filter them out...well, I draw conclusions.
Edit: its a paid for service hence the login required. Looks like its starts at $5/mo for 300 searches per month. How many search do you all average per month?
75 replies
I've replaced search engines with Wikipedia as the first step in looking for information. skimming an article is much faster than opening and closing a dozen sloppy s.e.o. 'd websites :p
I like the old ways because I also see/learn about unexpected things.
Let them waste money on generating an AI overview I'll happily skip on my way to the wikipedia link.
Wikipedia has an app. It's half of why I Google stuff. Usually a good place for quick answers, and an ok place to start digging in.
I'm half amazed at how bad Google has gotten, but that's probably just me remembering the early days of web searches.
That and TVTropes of course
That website was such a timesink. Amazing.
I have it disabled through an extension on my phone's Iceraven browser. Can't bother reading those. I need the Wikipedia/website info.
AI overview uses around 20 times more energy than regular searches.
It's absolutely batshit crazy that most search engine put it on by default just for it to summarize a wiki page 90% of the time. It's like we're collectively pushing for a mass extinction.
And had a useful answer about 1 in 20 times.
What's even crazier is for a lot of things we already had a quick result overview. Like if you wanted the definition of a word. For a long time it was the case that you could just "define socks" and you'd get a widget at the top with the definition. Why would you replace that with AI, and have an AI process that request every single time someone wants to define socks instead of just keeping the old widget!?
Indeed.
I'm pretty sure a good percentage of the searches are also just people typing the name of the website they want to access in the search bar. They don't want any information, just a link. All these searches now consume a massive amout of energy for basically nothing.
"Innovation" lol
I sure would hope that those AI overviews are cached. Querying an LLMs for every repeated search sounds very wasteful
If it didn't do the AInsearch repeatedly they couldn't use the billions of daily searches to artificially boost their AI engagement metrics and further inflate this bubble.
Me after losing 6 days of testing broken theories from an ai search when I could have just spent an hour watching a video that a human made that was fully functional from the start.
That said: No doubt that human spent way longer discovering the technique and then making the video: this should also be considered as a reality that creating something good takes time. And humans are still better at creating scalable content just give them the proper R&D time ffs.
LLMs fundamentally can't learn. They will continue to fail the same thing over and over until they guess and check into a response that appears coherent and correct.
All jokes aside, I usually do want to read the six Wikipedia pages because getting a one-sentence answer with no context is risky and not all that useful anyway. I need the nuance and background info for any question worth asking.
If you need nuance and background, wikipedia is NOT the place to get that. Maybe you can get some good links at the bottom of the page to real information, but that can be hit or miss also.
It's like getting the info from some rando at some party. It's to be lightly penciled in.
Do you really mean questions about things people might not know that yet have simple answers that don't need pages of context, are not worth asking?
Okay so how much glue to keep the cheese on my pizza?
4 seems like a good amount.
I would say that even simple answers at least deserve a source or citation. This is my biggest complaint with AI search, you get an "answer" with no qualification. The citations listed in the AI summary often don't contain the actual info being presented.
This question certainly wasn't worth asking.
AI has never been 100% accurate to the documents it pretends to reference, so the fact that anybody could even consider reading it like it's not pure gibberish is mind boggling to me.
I gave Claude a document to proofread today. It came back with an error. Clause said I used "told" instead of "tell." The error ended up being Claude's error, as I used "tell," and it was just confused.
This was the first time I ever used Claude. Just wanted to see what it was like. It did apologize, in fairness. But yeah, it's clearly fallible. I guess I don't understand it well enough to get how it read the document incorrectly. I'm baffled.
It's because it's a statistical model that guesses the next word. You asked it to read a document and formulate a response. So, based on responses in the dataset to the documents in the dataset, it chose the words and phrases which were most likely to be in a response.
Statistically, the most likely result the machine could come up with, was that you had used the wrong word. Factually, you did not.
When the old new ways become the old ways and i cant keep up with the new new ways
But check out this awesome ai recommendation I got when searching for Halloween 2026... (Fuck everything about this btw, no sarcasm)
Ooo, they're cramming in the ads now eh?
Fucking hell, lol
Lmao, I guess my internal ad filter is still working because I had to look at that twice to notice it and wondered if there was a new Halloween movie coming out this year or something. Agreed, fuck that shit.
Yesterday there was a weird play in the Thursday Night Football game where the Bears' quarterback handed the ball to an offensive lineman, who ran it in for a touchdown. The referees threw a flag and called the play back because it was an "illegal handoff", which I had never seen before.
I was texting my brother during the game and we both commented on how weird it was. My brother googled it and read the AI summary, which said the play was legal - citing a reddit thread from several years ago. He was like "Google says the play was legal - the refs got it wrong!"
I had to go to the NFL rulebook and dig through a bunch of sections before confirming that the refs were correct and you can't do that.
AI is absolute dogshit for anything that requires a bit of nuance. But it is really fast at serving up that dogshit, so a lot of lazy people would rather just accept a quick wrong answer than a slow correct one.
It would have been a legal play IF the lineman had checked in with the referee as an "eligible" receiver before the play. But evidently he screwed up and report in. It's not a common play in the NFL, but every team has such a play on hand.
So that's kind of a "it needs context" question.
Yeah exactly. The AI was just wrong and said it's legal as long as you hand it off behind the line of scrimmage. I actually still have the screenshot he sent. Notice the 2nd bullet point.
It's not just wrong. It's plausibly wrong, which is worse. It's the difference between giving someone a recipe for cookies where you fuck up the ratio of flour and sugar vs giving someone a recipe for cooking where you've jokingly put "50 pounds of gunpowder" in there as an ingredient. People know the obvious wrong thing is wrong. But they'll bake the fucked up ratio cookies and then just not know what went wrong.
May I suggest DuckDuckGo? They let you turn off the AI summary.
They even have a direct URL for it if you don't want to change the setting: https://noai.duckduckgo.com/
Better yet:
https://noai.duckduckgo.com/
and you can add it as a custom search for firefox/waterfox as:
https://noai.duckduckgo.com/?q=%25s
This is nice but I never figured out the search suggestions query string.
They also give you the option to (do their best at) blocking AI-generated search results. (Not perfect, of course, but it does a decent job of reducing how much AI slop you get in your search results.)
noai.duckduckgo.com
The funniest part is I've gotten better answers from duckeuckgo and Bing's AI back in the day because they both didn't treat reddit as a reputable source of info lmao.
Dunno how the scene is now, but I wouldn't be surprised if its the same.
My go to. I used to use Google for harder to search stuff, but these days Google isnt worth the try anymore.
I refuse to use AI but my boss only uses it. He has not found a correct answer to a query for 6 months.
This grew from ‘I don’t want Wikipedia, I’ll pay $150 a month to Encyclopaedia Brittannica for a set of 60 books the way God intended…’
Encarta FTW
There's a site for that..let me remember the name.. Wikipedia! There it is.
This is the way
You may either read all of them up front, or read all of them when verifying the validity of the ai summary. Your choice.
THIS! I recall somebody doing contract review and having AI summarize the contracts just to see how well they did... The AI left out some really important stuff and also occasionally hallucinated stuff that wasn't there. Summarizing a contract should be pretty black-n-white -- if AI can't handle that properly, I'm not going to trust its summary of something more complex.
6? That's rookie numbers man, you gotta pump those numbers up.
The Wikipedia marathon doesn't stop til you get to Hitler.
That's right. I didn't put the question in the search bar for you to tell me the answer. That question was the first shoveling of dirt in a 6 hour rabbit hole that I was fixing to explore.
Yep. I still have yet to see an actual use for AI. People are like “but summaries!” Summarize this: summaries are only useful if the information being summarized doesn’t matter, which to me means you could just skip the entire thing.
This isn’t a high school test, this is real life. Why are you trying to use SparkNotes?
It's great for when you want made-up bullshit.
Fake product reviews, for example. AI is fantastic for that. Want to buff up your shitty product's Amazon reviews? AI can generate the text for those reviews perfectly well.
As someone with ADHD, I disagree. Sometimes the information is important, and I don't want to have to dig through a wall of text to find it. One thing that annoys me about normies when they write, is that they tend to blab on forever on things that don't matter instead of getting straight to the point.
If people wouldn't use so many unnecessary words, I wouldn't need a god damn LLM to summarize their ramblings into something my brain can actually parse.
They are great to tell you the words you need to search for something; they are great for analyzing "vibes" out of texts; they are reasonably good to generate you a link to some site that answers a hard to find question, traditional search would be better if there were any one around that isn't completely broken.
I hate it when i need to google programming stuff because Google Gemini LOVES to make stuff up and takes up half a page trying to answer
stop googling, start ducking.
is "ducking" an actual term now? Come to think of it, ducking doesn't sound any sillier than googling
Don't you dare duck with me!
I know the rankings change like every month, but isn't Gemini one of the "dumber" ones? I've only tried it a few times and got annoyed, but it's been a while.
Oh it is, it has the same issues as the first few iterations of Chat GPT if you ever used AI in the early days
It's certainly no leader for software engineering tasks.
If you use Google just type -AI after your search and it removes the AI.
join the elite: kagi.com
Join the elite: create an account and pay $120/year to a private American company for search results.
Don’t worry, they’re also super AI-pilled and known for randomly cancelling projects. But I’m sure the search is great or whatever
That's still full of AI, the filter quality you can see for yourself without an account for a limited amount of searches. If I search an image and the URLs on slop I see in the results with the slop filter enabled that literally contains keywords related to AI slop and KAGI doesn't filter them out...well, I draw conclusions.
A search engine you have to sign up for?
Edit: its a paid for service hence the login required. Looks like its starts at $5/mo for 300 searches per month. How many search do you all average per month?
And pay, i'd not mind a sign up but i'd rather not pay, especially if i cannot even have the option to self host
once you know, you can never go back
Hey, free socks. Actually, I think that promo is over.
I tried them and looked into them. They do something, but nothing like the curated searches they try to sell you on.
…and then have an AI summarize those six Wikipedia articles because you can't find the answer you're looking for in them.