I’ve experienced the opposite. I get good summaries with links for original sources when something seeems off, which it almost never does. I also like that I can glance to see where the data came from.

Several months ago, I was getting hallucinations and silly answers, like Jaco Pastoreous being the bass player for Metallica.

but lately it’s been great. Especially for technical syntax stuff, like asking for syntax on a jq command, or sal syntax, or something trivial like movie casts members….

Edit: I also assume there’s now some pre-processing filter to retrieve answers from a cache or something, because I’m getting pretty long answers very quickly….

While I agree in principle, I think it was inevitable. People been gaming SEO for so long that even with judicious use of search operators, it was getting harder to find the things I wanted (probably drowned in a sea of spam in such a way as to fall outside of the optimized search index). It _looks like_ the AI overview does not have this problem (yet...).

>All the "AI search" I've tried is far less relevant than a proper search, using search engines without it seems much faster and more relevant.

I've had the opposite experience with Perplexity Pro. Yesterday it took just under nine minutes to deliver a detailed and accurate answer as to the nearest locations where I can buy Meta Ray-Ban Display glasses (at least 3 hours drive away FWIW) that day.

In the past several months I've spent countless hours using search engines with dismal results that weren't close to that.

It is totally dependant on context. Ask about product purchases/prices and it will give you great results. Ask for basic data and it will quote wikipedia or government websites all day. Ask for a coding solution and it will spit out answers ripped from the corners of github. But i asked google about hydrogen-3 recently and got back a result saying that we were mining it on the moon. An entire space economy was invented because AI cannot tell scifi from reality.

The real test: dont test AI by asking it questions you already know. Ask it a proper question and then go out and find the answer yourself. That is where AI really falls down, the non-trivial stuff.

I've noticed something of a hybrid of what you posted and what one of your responders below posted.

For average results, that can have an acceptable answer that tends to the mean or societal mainstream views along some subject, the google answer works for me. For others, which are well unfortunately most of my queries these days, the google answer is brainwashing and propaganda. It feels little more than a continuation of the SARS-COV-2 era fact checkers which were actually failed attempts at propaganda machines for enforcing certain views as true while others were cited as "evidence does not suggest" (hand selecting the evidence of course means this is nothing more than confirmation bias).

I think the AI technology used in most LLM models is an acceleration of this- it will tout mainstream views or hand-picked views as truth while calling others false. And without an ability to search any more for opposing views, you get left with a really warped view of reality that might work for most. Though not for everyone.

Regarding a swiffer: it's a zero sum game. Where did the materials come from? The earth. Where did they return after use? The earth. The only thing changed was the energy used to create and transport them.

If conserving energy was important, none of these AI datacenters would be built, none of our items would be shipped across the planet from Asia. They would conserve energy and import raw materials only.