We must have wildly different workflows/ways to interact with the web.
Search is always faster than asking an LLM if I have a general idea of what I am looking for. I may consult an LLM if I want to compare things or kick off deep research, but most of the time I find myself having to go back and forth with it and correcting assumptions it made.
According to my Kagi stats, I am averaging around 3k searches per month.
I can’t help but feel that you are really missing out on a lot of results when just relying on LLMs for search.
I guess both my job and my personal life do involve a lot of searching, though i’d say that sometimes (a lot of the time) its also my ADHD getting the best of me.
Looking through some of my history of today:
- “github rate-limits”
- “oriental hornet”
- “riva 88 florida”
- “logistic map”
- “zeiss euv mirror”
- “authentik helm”
Just to name a few…
For a lot of these, LLMs would slow me down significantly. Most of the time I already know exactly what im looking for.
I'm with the GP. To me LLMs are just better search engines. In the most literal sense. They have their own index and can generate links if you want them.
By “their own index” i assume you mean training data?
If so, thats a big part of the issue for me. As an example, if I ask an LLM for some part of the Zig stdlib, I will get an incorrect answer 10/10 times because it will refuse to look up the latest documentation.
No, I mean they're constantly crawling the web. That gets fed into the next model but it's also searchable by the current one. For example, if you ask chatgpt what the latest shooting at the white house was it will tell you about the one from two days ago.
ChatGPT has effectively assimilated Google search. You can tell it to look up the latest version of the zig docs and do whatever.
Sure, yes. But that has been quite flakey for me as well.
Regardless, if we stick with the example of the Zig docs, using the search and then opening the actual docs is a much better and faster experience for me. I get the context I need, I don’t get the verbose llm output that packages it and I get there faster.
Mind you, I use AI a lot and have subscriptions to all of the major models, but this is just not a use-case I find useful in my workflow.
I also keep trying to use it this way by looking for things such as decent libraries or frameworks for something new im trying out. I’d say that at least 50% of the time I end up searching for myself and finding popular ones that were completely ignored by whatever LLM I searched with. Always a frustrating experience…
If you end a search query with a question mark, kagi answers with their version of search overview. But with a quality closer to asking an agent with access to your search results. It's great for one-off queries
Search is always faster than asking an LLM if I have a general idea of what I am looking for. I may consult an LLM if I want to compare things or kick off deep research, but most of the time I find myself having to go back and forth with it and correcting assumptions it made.
According to my Kagi stats, I am averaging around 3k searches per month.
I can’t help but feel that you are really missing out on a lot of results when just relying on LLMs for search.