I was watching the RFK Jr questioning today and when Bernie was talking about healthcare and wages I felt he was the only one who gave a real damn. I also thought “Wow he’s kinda old” so I asked my phone how old he actually was. Gemini however, wouldnt answer a simple, factual question about him. What the hell? (The answer is 83 years old btw, good luck america)
Why would you use a chatbot to attempt to obtain factual information?
Especially something so trivial. If you use it to learn about some larger conflict or something, fine (though don’t expect accuracy). If you’re using for age, which has been trivial to find with a quick search for at least a decade, something has gone wrong with you. It’s the higher effort option for a worse result.
To see if it can do it and how accurate its general knowledge is compared to the real data. A locally hosted LLM doesnt leak private data to the internet.
Most webpages and reddit post in search results are themselves full of LLM generated slop now. At this stage of the internet if your gonna consume slop one way or the other it might as well be on your own terms by self hosting an open weights open license LLM that can directly retrieve information from fact databases like wolframalpha, Wikipedia, world factbook, ect through RAG. Its never going to be perfect but its getting better every year.
I mostly can’t understand why people are so into “LLMs as a substitute for Web search”, though there are a bunch of generative AI applications that I do think are great. I eventually realized that for people who want to use their cell phone via voice, LLM queries can be done without hands or eyes getting involved. Web searches cannot.
Because web search was intentionally hobbled by Google so people are pushed to make more searches and see more ads.
I still use web search all the time, I just don’t use Google. There are great alternatives:
DDG and Ecosia are proxies for Bing. I didn’t check, but I’m guessing the others are too. Most “independent” search engines are.
The major exception is Startpage, which is a proxy for Google.
Marginalia is not.
Marginalia and Million Short have their own index as far as I know. Fireball is another one being independent from both Google and Bing.
startpage is a proxy for google?
That one’s pretty obvious. From their main page:
Do any of them actually work? As in, you search something and it gives you relevant results to the whole thing you typed in?
I’ve used DuckDuckGo for a long time, so I would say yes. But the best way to figure that out is just to try it for a while. There is literally nothing to lose.
I have been using it for about 7 years and its just as shit.
It’s gotten really bad lately. Changing search terms in my query to the point where it’s searching for something completely different.
Also, it’s extremely us centric. It will frequently return results from shops that do not exist where I live in favour of american ones
It’s always going to depend on what you’re searching for. I just tried searching for home coffee roasting on Swiss Cows and all of the results were legit, no crappy spam sites.
Marginalia is great for finding obscure sites but many normal sites don’t show up there. Million Short is a similar idea but with a different approach to achieving it.
The problem of search is actually extremely hard because there are millions of scam and spam sites out there that are full of ads and either AI slop or literally stolen content from other popular sites. Somehow these sites need to be blocked in order to give good results. It’s a never-ending, always-evolving battle, just like blocking spam in email (I still have to check my spam folder all the time because legit emails end up flagged as spam).
Web search by voice was a solved problem in my recent memory. Then it got shitty again
google literally has a voice button.
The problem isn’t conducting the search with voice, it’s receiving any actual information back. A few years ago I would ask a question and receive an answer based off the top few results, and if it couldn’t scrape something together it would just give me the results instead.
I haven’t used voice search in a while because of the issues that started to arise, but I have less fond memories of “hey Siri, answer this.” And then having to go find my phone anyway to Google it because she was useless
Would saying “Gemini, open the Wikipedia page for Bernie Sanders and read me the age it says he is”, for example, suffice as a voice input that both bypasses subject limitations and evades AI bullshitting?
Gemini refuses to answer
Copilot seems to know the current date and calculates the age from that
ChatGPT is clueless
Idk if it bypasses limitations, you can try. As for bullshiting, no. The AI almost certainly does not have the ability to go and open a webpage. If it was trained on wikipedia, it may or may not give you the age listed at the time of it’s training. If not, it will likely take a different source and pretend it is from wikipedia. Either way, it will likely bullshit you about doing what you asked while giving you outdated/missourced information.
Now the number may be correct, I imagine Bernies real age is readily available, but it will confidently lie about how it got the information.
To be honest, that seems like it should be the one thing they are reliably good at. It requires just looking up info on their database, with no manipulation.
Obviously that’s not the case, but that’s just because currently LLMs are a grift to milk billions from corporations by using the buzzwords that corporate middle management relies on to make it seem like they are doing any work. Relying on modern corporate FOMO to get them to buy a terrible product that they absolutely don’t need at exorbitant contract prices just to say they’re using the “latest and greatest” technology.
To be honest, that seems like it should be the one thing they are reliably good at. It requires just looking up info on their database, with no manipulation.
That’s not how they are designed at all. LLMs are just text predictors. If the user inputs something like “A B C D E F” then the next most likely word would be “G”.
Companies like OpenAI will try to add context to make things seem smarter, like prime it with the current date so it won’t just respond with some date it was trained on, or look for info on specific people or whatnot, but at its core, they are just really big auto fill text predictors.
Yeah, I still struggle to see the appeal of Chatbot LLMs. So it’s like a search engine, but you can’t see it’s sources, and sometimes it ‘hallucinates’ and gives straight up incorrect information. My favorite was a few months ago I was searching Google for why my cat was chewing on plastic. Like halfway through the AI response at the top of the results it started going on a tangent about how your cat may be bored and enjoys to watch you shop, lol
So basically it makes it easier to get a quick result if you’re not able to quickly and correctly parse through Google results… But the answer you get may be anywhere from zero to a hundred percent correct. And you don’t really get double check the sources without further questioning the chat bot. Oh and LLM AI models have been shown to intentionally lie and mislead when confronted with inaccuracies they’ve given.
I think that one major application is to avoid having humans on support sites. Some people aren’t willing or able or something to search a site for information, but they can ask human-language questions. I’ve seen a ton of companies with AI-driven support chatbots.
There’s sexy chatbots. What I’ve seen of them hasn’t really impressed me, but you don’t always need an amazing performance to keep an aroused human happy. I do remember, back when I was much younger, trying to gently tell a friend who had spent multiple days chatting with “the sysadmin’s sister” on a BBS that he’d been talking to a chatbot – and that’s a lot simpler than current systems. There’s probably real demand, though I think that this is going to become commodified pretty quickly.
There’s the “works well with voice control” aspect that I mentioned above. That’s a real thing today, especially when, say, driving a car.
It’s just not – certainly not in 2025 – a general replacement for Web search for me.
I can also imagine some ways to improve it down the line. Like, okay, one obvious point that you raise is that if a human can judge the reliability of information on a website, that human having access to the website is useful. I feel like I’ve got pretty good heuristics for that. Not perfect – I certainly can get stuff wrong – but probably better than current LLMs do.
But…a number of people must be really appallingly awful at this. People would not be watching conspiracy theory material on wacky websites if they had a great ability to evaluate it. It might be possible to have a bot that has solid-enough heuristics that it filters out or deprioritizes sources based on reliability. A big part of what Web search does today is to do that – it wants to get a relevant result to you in the top few results, and filter out the dreck. I bet that there’s a lot of room to improve on that. Like, say I’m viewing a page of forum text. Google’s PageRank or similar can’t treat different content on the page as having different reliability, because it can only choose to send you to the page or not at some priority. But an AI training system can, say, profile individual users for reliability on a forum, and get a finer-grained response to a user. Maybe a Reddit thread has material from User A who the ranking algorithm doesn’t consider reliable, and User B who it does.