LLMs and their precursors were always how search engines compiled and ranked their search results. Google has been using some form of AI to build up its base of information for decades.
The absurd turn of the last decade was Google C-level pivot from engagement through useful results to engagement through bad results that force you to click around on the site longer. “How do we get more ad interaction?” became the question that dominated the development of their technology.
given that by LLMs, people generally refer to the transformer architecture, it definitely wasn’t how google always did search. they famously started out by using pagerank. now, back then this was definitely seen as AI, but what gets called ai by the general public has changed over the last couple of decades. now people use it almost exclusively to refer to deep learning models or even just llms.
so yes, in a strict sense google has always used ai for search. but when the general public says AI now, they usually mean llms specifically and using that to replace actually going to the source websites is new
Pagerank was built on the Markov Model, which became the core of modern LLM graph design.
It was a precursor technology, not an independent system, which used Hyperlinks rather than word tokens to form associations.
they usually mean llms specifically and using that to replace actually going to the source websites is new
So much of the complaint around AI isn’t even the LLM part. It’s the weighting and censorship of certain sources and data, convinced with the masking of results behind natural language output templates.
You had the first problem under Pagerank already (Google Bombing and “shadow banning” of sources the company didn’t find reliable). The second is entirely a product of the interface and has nothing to do with the heuristics of the program.
sure both pagerank and llms can modelled as marlov chains, but the working and output of these models is very different. it’s disingenuous to state that google used llms and their precursors since the beginning. their precursors maybe but they are very different beasts.
as for the complaints, for me a big problem of their use in search engines is that they often hide the sources of the summary they provide and disincentivize people from clicking through to the source sites. This is very much a new problem that didn’t exist before google included output from generative models
Impossible. The problem is not in the search process, but in data availability. LLM gives you a concrete concise answer to your question while the search engine gives you dead links or spam.
A concrete, concise answer that’s often confidently incorrect.
I’ve had MS Copilot first tell me to use a Microsoft product that was retired nearly a decade ago, and when called out it told me to use one that didn’t exist.
Also, the garbage web results are because of LLMs.
LLM gives you a concrete concise answer to your question while the search engine gives you dead links or spam.
LLM also very confidently gives you a wrong answer way too often. Just over the weekend I tinkered with diy-home automation gadget and fed a bunch of documents to duck.ai in hope that it could dig up protocol and electrical behavior of HAN port on our electric meter. The documentation is there, but it’s recursively referenced to a crapload of different standards so the actual answer is pretty labour intensive to dig up manually.
LLM spat out an answer which seemed to make sense, but the solution it offered didn’t work and when I asked for sources for spesific facts it just said that it doesn’t have any and that it shouldn’t have recommended such a solution in the first place, with apologies of course.
I’d much prefer the old school way of searching things, but that doesn’t really exist anymore either. Search engines, google specially, are far less useful than they used to be. Part of the problem is LLM slop filling the whole internet but somehow I feel that’s not the only reason.
Here’s an example of a time where a single letter spelling error affected the outcome even though Google knew it was a typo. Yea, LLMs will give an answer but anecdotally it’s correct less than half the time on any question of substance and at least 10%+ miss rate on simple questions.
Ditching Google is good, but not for AI. We should take the ditching part and redirect the young people to use other search engines.
LLMs and their precursors were always how search engines compiled and ranked their search results. Google has been using some form of AI to build up its base of information for decades.
The absurd turn of the last decade was Google C-level pivot from engagement through useful results to engagement through bad results that force you to click around on the site longer. “How do we get more ad interaction?” became the question that dominated the development of their technology.
That’s what enahittified their website. Not “AI”.
given that by LLMs, people generally refer to the transformer architecture, it definitely wasn’t how google always did search. they famously started out by using pagerank. now, back then this was definitely seen as AI, but what gets called ai by the general public has changed over the last couple of decades. now people use it almost exclusively to refer to deep learning models or even just llms.
so yes, in a strict sense google has always used ai for search. but when the general public says AI now, they usually mean llms specifically and using that to replace actually going to the source websites is new
Pagerank was built on the Markov Model, which became the core of modern LLM graph design.
It was a precursor technology, not an independent system, which used Hyperlinks rather than word tokens to form associations.
So much of the complaint around AI isn’t even the LLM part. It’s the weighting and censorship of certain sources and data, convinced with the masking of results behind natural language output templates.
You had the first problem under Pagerank already (Google Bombing and “shadow banning” of sources the company didn’t find reliable). The second is entirely a product of the interface and has nothing to do with the heuristics of the program.
sure both pagerank and llms can modelled as marlov chains, but the working and output of these models is very different. it’s disingenuous to state that google used llms and their precursors since the beginning. their precursors maybe but they are very different beasts.
as for the complaints, for me a big problem of their use in search engines is that they often hide the sources of the summary they provide and disincentivize people from clicking through to the source sites. This is very much a new problem that didn’t exist before google included output from generative models
Search is probably one of the few “ok” uses for LLMs if you find a way to reduce the carbon footprint and train on ethical material
Which, up until maybe eight or nine years ago, Google was pretty good at.
But then we went into this insane build-out cycle, where you were seemingly incentivized to be extra sloppy with every aspect of the business.
I’ve been enjoying using https://marginalia-search.com/ for various topics.
Impossible. The problem is not in the search process, but in data availability. LLM gives you a concrete concise answer to your question while the search engine gives you dead links or spam.
A concrete, concise answer that’s often confidently incorrect.
I’ve had MS Copilot first tell me to use a Microsoft product that was retired nearly a decade ago, and when called out it told me to use one that didn’t exist.
Also, the garbage web results are because of LLMs.
LLM also very confidently gives you a wrong answer way too often. Just over the weekend I tinkered with diy-home automation gadget and fed a bunch of documents to duck.ai in hope that it could dig up protocol and electrical behavior of HAN port on our electric meter. The documentation is there, but it’s recursively referenced to a crapload of different standards so the actual answer is pretty labour intensive to dig up manually.
LLM spat out an answer which seemed to make sense, but the solution it offered didn’t work and when I asked for sources for spesific facts it just said that it doesn’t have any and that it shouldn’t have recommended such a solution in the first place, with apologies of course.
I’d much prefer the old school way of searching things, but that doesn’t really exist anymore either. Search engines, google specially, are far less useful than they used to be. Part of the problem is LLM slop filling the whole internet but somehow I feel that’s not the only reason.
LLMs give you sycophantic bullshit. Tell it it’s wrong even if the response is correct and see what it does.
That may be true for basic, simple questions, but it is not true for medium or hard questions.
They will give an answer, yes, but often it’s not answering what you asked, or giving the wrong answer.
Here’s an example of a time where a single letter spelling error affected the outcome even though Google knew it was a typo. Yea, LLMs will give an answer but anecdotally it’s correct less than half the time on any question of substance and at least 10%+ miss rate on simple questions.
Then you use a better search engine, not flock to an AI that has no real concept of what’s true or false so just feeds you whatever looks right