Respectfully it's up to the maintainer who decides where they think it shoudl go.
After I read the same I thought, ok, that's what _they_ wanted to do.
Me? I'd love to see what could be rewritten to be 100% future proofed for as many years as possible. But see what that did? I have to share, and see if anyone else already thought of this (I doubt I'm the first ever to have this thought), and maybe, I'll be lucky enough to benefit from someone exploring it already, before I decide to put on a jedi robe and wave my hands to use the force to move AI around.
It doesn't /inherently/ return better search results, but Google is many years into an intensive project to make their search results so bad that even a magic eight ball, let alone a moderately advanced AI, can do better.
Google's organic search results are ranked by quality (including how many people click on them).
AI, generally thinks the best result is what's most common, or average, resulting in decent results, but not the best for what you're looking for.
We notice it when we're looking for a level of specificity google can hit, compared to an ai model. It's why Google's AI search can hit a bit differen too.
The example that comes to top of mind is a search that is not made for search rarely can compare to something as well as Perplexity.
> Google's organic search results are ranked by quality
What does this even mean? They're ranked by metrics that are assumed to be a proxy for quality, but (a) the accuracy of those proxies varies widely and (b) the ranks are relative; if a result is no longer in the response pool the fact that it would have been higher quality is invisible. Consider the year where anyone who's profile had been tagged as a software developer couldn't find any information on tost rings for sailplane towing because the "autocorrection" to toString was so strong that no mechanism besides logging out gave any relevant results; there's always crap like this going on.
When I'm looking for info that requires careful parsing of a source site, then LLM is faster. If it's something like looking up the hours of a business googling that local business is better imo. Result is fast you get to see the business picture etc
Search has moved on from tf.idf + pagerank at this point. And absolutely they do use things like encoders to produce embeddings that are used in retrieval.
Google ranked the best results, not the most commonly occurring ones.
AI is by default returning the highest statistical percentage of what it thinks you are looking for in a response, which may or may not be a search result.
A fun example is having the AI run a search, and comparing it to perplexity, or even gemini a little. OpenAI is improving, and Claude is pretty consistently not made for comprehensive search.
After I read the same I thought, ok, that's what _they_ wanted to do.
Me? I'd love to see what could be rewritten to be 100% future proofed for as many years as possible. But see what that did? I have to share, and see if anyone else already thought of this (I doubt I'm the first ever to have this thought), and maybe, I'll be lucky enough to benefit from someone exploring it already, before I decide to put on a jedi robe and wave my hands to use the force to move AI around.
reply