[imzadi]

LLM's are just summarizing search results and not always well. This is like asking if academic papers are relevant when college essays exists, imo. I would much rather read the original source of the material than the summarized (potentially badly) LLM result.

[gf000]

One caveat is that many existing search engines are actively worse than what they had before.

Exact keyword matching had its use, but now they do some kind of vector search but without the text understanding of LLMs. Making anything more complex impossible to find without actually going to an LLM that understands what I'm actually going for and can help out.

Just recently I wanted to find out the mill build tool's special syntax to have the JVM version used for the build tool itself be the system version (on nixos, downloading a JVM won't work, but I can trivially pin a jdk version myself. Unfortunately mill will refuse to even start without this special line).

My Google/kagi fu absolutely failed, yet Claude managed to find me the exact syntax in a single prompt. (Do give it a try if you want to)

[hs86]

Compared to Claude and Gemini, I think ChatGPT is really more thorough when it comes to web searches and providing sources. The others usually answer in a short paragraph and give me maybe one or two sources, while ChatGPT thinks for a minute and presents me a little mini report with 10-20 sources to look further into. This is honestly the reason why I stopped paying for Kagi, but I'm not sure how worse this will get when ads become commonplace in ChatGPT.

[hyperionultra]

Convenience is a massive leverage here. And most people trust LLM de facto. If a tool is good enough for 85% of audience - anything else is irrelevant.

[fassssst]

Well you could ask it to only reply with snippets from its sources.

[threetonesun]

So a slower search engine with fewer results?

[hilariously]

Delivery food is not faster or better in a lot of places.

[dexterdog]

Delivery food is worse everywhere I've ever had it.