Comment by kdheiwns | Hacker News Reader

Comment by kdheiwns | Hacker News Reader

Parent Full thread

I used to be able to google a question like that and get an accurate answer within the top 3 results nearly every time about 20 years ago. Then it got worse and worse and became pretty much completely useless about 10 years ago.

Now AI will give me a confident answer that is outright wrong 20% of the time or kind of right but not really 30% of the time. So now I ask something using an AI chatbot and carefully word it so as to have it not get off topic and focus on what I actually want to know, wait 30 seconds for its long ass answer to finish, skim it for the relevant parts, then google the answer and try to see where the AI sourced its answer from and determine whether it misinterpreted/mixed up results or it's accurate. What used to be a 10 second google search is now a 2-3 minute exercise.

I can see very much how people say AI has somehow led to productivity losses. It's shit like this, and it floods the internet and makes real info harder to find, making this cycle worse and worse and take more and more time for basic stuff.

Same experience here. I have fond memories of “google code”, a search engine for code databases which was exceptionally good for finding literal quotes.

The more mainstream a subject is, the lower the incidence of hallucinations. With google search, the mantra “I can’t be the first with this problem/question” almost always proves to be right.

I’m in the process of restoring a piece of vintage electronics and everytime I ask gemini (fast or thinking) for help I’m getting sent down an irrelevant rabbit hole. It’s taking info from service manuals of other equipment with a similar product number, misinterpreting diagrams, getting electrical workings wrong.

These things aren’t AI. AI can extract certainty from uncertain data. LLMs take data and turn it into garbage.

Web scraping for LLMs has almost completely ruined the search experience. In the past I could search for simple questions, and quickly get an answer without even having to click through to the links.

This was horrible for web traffic, but the utility level was off the charts. It was possible to get accurate results in milliseconds. It was faster than using an LLM.

Now sites put almost no info in the search result headers, to get people to click through. I think this will work on some users, but most will start using LLMs as search by default.

Search engines have gotten so bad that I almost feel forced to try running SearXNG or some other search engine locally. Its a pain to set up, but degooglefication is always worth it.

Is SearXNG a nice experience? I need something that specifically excludes AI slop affiliate marketing listicles.

I don't know if content other than that exists out there any more. Still, I'd rather have no answer than a "top ten $item" list just describing the first page of Amazon results.

Once its dialed in, its as good as Kagi. But I still haven't tried to integrate it into all my devices outside of my computer. I still need to try it with tailscale.

I also feel like a really good local LLM can be better. I'd love to get something set up where my searches get sent to the local LLM first, and then to Kagi\SearXNG if it needs more information.

Now Google has an AI answer at the top with links to sources. This streamlines the process.

Reply on news.ycombinator.com