I keep having to repeat this, but the conversation does keep going on a loop: LLMs aren’t entirely useless and they’re not search engines. You shouldn’t ask it any questions you don’t already know the answer to (or have the tools to verify, at least).
Or if you’re fine with non-factual answers. I’ve used chatgpt various times for different kinds of writing, and it’s great for that. It can give you ideas, it can rephrase, it can generate lists, it can help you find the word you’re trying to think of (usually).
But it’s not magic. It’s a text generator on steroids.
Sure! Used as… you know, what it is, there’s a lot of fun/useful stuff you can do. It’s just both AIbro shills and people who have decided to make hating on this tech a core part of their personality have misrepresented that.
It’s indeed very, very good text generation/text parsing. It is not a search engine, the signularity, Skynet or a replacement for human labor in the vast majority of use cases.
“Intended” is a weird choice there. Certainly the people selling them are selling them as search engines, even though they aren’t one.
On DDG’s implementation, though, you’re just wrong. The search engine is still the search engine. They are using an LLM as a summary of the results. Which is also a bad implementation, because it will do a bad job at something you can do by just… looking down. But, crucially, the LLM is neither doing the searching nor generating the results themselves.
honestly LLMs are about a thousand times more useful than Google at this point. Every week i try googling and get nothing but spam results.
for example just yesterday i was searching for how to reclaim some wasted space on one of my devices. so i searched on Google and tried 8 different pages that were ad-riddled hell holes.
i gave up and spent 10 seconds with an LLM and got the answer i needed. i will admit that i had to tell it to quit bullshitting me at one point but i got what i needed. and no ads.
Well, you shouldn’t be using Google Search, but that’s a completely different conversation and the answer shouldn’t (can’t) be “let’s just use LLMs, then”.
bing or duck duck go, too. i just say googling because it sounds stupid as shit to say anything else. DDG is my default search engine. kagi isn’t much better, and comes with its own issues
Yeah. Everyone forgot the second half of “Trust, but Verify”. If I ask an LLM a question, I’m only doing it because I’m not 100% sure how to look up the info. Once it gives me the answer, I’m checking that answer with sources because it has given me a better ability to find what I was looking for. Trusting an LLM blindly is just as bad as going on Facebook for healthcare advice.
I find LLMs very useful for setting up tech stuff. “How do I xyz in docker?” It does a great job of boiling together several disjointed How Tos that don’t quite get me there into one actually usable one. I use it when googling and following articles isn’t getting me anywhere, and it’s often saved so much time.
They are also amazing at generating configuration that’s subtly wrong.
For example, if the bad LLM generated configurations I caught during pull requests reviews are any example, there are plenty of people with less experienced teams running broken kubernetes deployments.
Now, to be fair, inexperienced people would make similar mistakes, but inexperienced people are capable of learning with their mistakes.
Yep. Or because you can recognize the answer but can’t remember it off the top of my head. Or to check for errors on a piece of text or code or a translation, or…
It’s not “trust but verify”, which I hate as a concept. It’s just what the tech can and cannot do. It’s not a search engine finding matches to a query inside a large set of content. It’s a stochastic text generator giving you the most likely follow up based on its training dataset. It’s very good autocorrect, not mediocre search.
LLM is a LLM. LLM is a transformer model generating likely output from a dataset.
I hate all this analogy stuff people keep resorting to. The thing does what it does, and trying to understand what it does by analogy is being used disingenuously to push all sort of misinformation-filled agendas.
It’s not about “trust”, it’s about how the output you’re being given is generated, and so what types of outputs are useful on what applications.
The answer is fairly narrow, particularly compared to how it’s being marketed. It absolutely, 100% isn’t a search engine, though. And even when plugged into a search engine and acting as a summarization engine it’s actually pretty terrible and very likely to distort an output that anybody who has been near a computer in the past thirty years can parse faster at a glance.
That is exactly the point, LLM aim to simulate the chaotic best guess flow of the human mind, to be conscious and at least present the appearance of thinking and from that to access and process facts but not be a repository of facts in themselves. The accusation here that the model constructed a fact and then built on it is missing the point, this is exactly the way organic minds work. Human memory is constantly reworked and altered based on fresh information and simple musings and the new memory taken as factual even while it is in large part fabricated, and to an increasing extent over time. Many of our memories of past events bear only cursory fidelity to the actual details of the events themselves to the point that they could be defined as imagined. We still take these imagined memories as real and act upon them exactly as has been done here by the AI model.
As below, stop with the analogies. No, that’s not “the chaotic best guess flow of a human mind”, that’s a whole bunch of tensor math generating likely chains of tokens. Those two things aren’t the same thing.
They aren’t the same thing in the strict sense, but they’re also not the same thing in practical terms at the end user level. If I ask a friend if they remember some half-forgotten factoid they can tell me not just if they do remember, but also how well they remember, how sure they are and why they know it. No LLM can do that, because LLMs know as little about themselves as about anything else. Which is nothing, because they’re LLMs, not people.
I keep having to repeat this, but the conversation does keep going on a loop: LLMs aren’t entirely useless and they’re not search engines. You shouldn’t ask it any questions you don’t already know the answer to (or have the tools to verify, at least).
Or if you’re fine with non-factual answers. I’ve used chatgpt various times for different kinds of writing, and it’s great for that. It can give you ideas, it can rephrase, it can generate lists, it can help you find the word you’re trying to think of (usually).
But it’s not magic. It’s a text generator on steroids.
Sure! Used as… you know, what it is, there’s a lot of fun/useful stuff you can do. It’s just both AIbro shills and people who have decided to make hating on this tech a core part of their personality have misrepresented that.
It’s indeed very, very good text generation/text parsing. It is not a search engine, the signularity, Skynet or a replacement for human labor in the vast majority of use cases.
I had to tell DDG to not give me an AI summary of my search, so its clearly intended to be used as a search engine.
“Intended” is a weird choice there. Certainly the people selling them are selling them as search engines, even though they aren’t one.
On DDG’s implementation, though, you’re just wrong. The search engine is still the search engine. They are using an LLM as a summary of the results. Which is also a bad implementation, because it will do a bad job at something you can do by just… looking down. But, crucially, the LLM is neither doing the searching nor generating the results themselves.
honestly LLMs are about a thousand times more useful than Google at this point. Every week i try googling and get nothing but spam results.
for example just yesterday i was searching for how to reclaim some wasted space on one of my devices. so i searched on Google and tried 8 different pages that were ad-riddled hell holes.
i gave up and spent 10 seconds with an LLM and got the answer i needed. i will admit that i had to tell it to quit bullshitting me at one point but i got what i needed. and no ads.
Well, you shouldn’t be using Google Search, but that’s a completely different conversation and the answer shouldn’t (can’t) be “let’s just use LLMs, then”.
bing or duck duck go, too. i just say googling because it sounds stupid as shit to say anything else. DDG is my default search engine. kagi isn’t much better, and comes with its own issues
Yeah. Everyone forgot the second half of “Trust, but Verify”. If I ask an LLM a question, I’m only doing it because I’m not 100% sure how to look up the info. Once it gives me the answer, I’m checking that answer with sources because it has given me a better ability to find what I was looking for. Trusting an LLM blindly is just as bad as going on Facebook for healthcare advice.
I thought it was “butt verify” whoops
I find LLMs very useful for setting up tech stuff. “How do I xyz in docker?” It does a great job of boiling together several disjointed How Tos that don’t quite get me there into one actually usable one. I use it when googling and following articles isn’t getting me anywhere, and it’s often saved so much time.
They are also amazing at generating configuration that’s subtly wrong.
For example, if the bad LLM generated configurations I caught during pull requests reviews are any example, there are plenty of people with less experienced teams running broken kubernetes deployments.
Now, to be fair, inexperienced people would make similar mistakes, but inexperienced people are capable of learning with their mistakes.
Yep. Or because you can recognize the answer but can’t remember it off the top of my head. Or to check for errors on a piece of text or code or a translation, or…
It’s not “trust but verify”, which I hate as a concept. It’s just what the tech can and cannot do. It’s not a search engine finding matches to a query inside a large set of content. It’s a stochastic text generator giving you the most likely follow up based on its training dataset. It’s very good autocorrect, not mediocre search.
LLM is a random person in the internet, or the first link on a search.
If you wouldn’t blandly trust them, don’t trust it.
LLM is a LLM. LLM is a transformer model generating likely output from a dataset.
I hate all this analogy stuff people keep resorting to. The thing does what it does, and trying to understand what it does by analogy is being used disingenuously to push all sort of misinformation-filled agendas.
It’s not about “trust”, it’s about how the output you’re being given is generated, and so what types of outputs are useful on what applications.
The answer is fairly narrow, particularly compared to how it’s being marketed. It absolutely, 100% isn’t a search engine, though. And even when plugged into a search engine and acting as a summarization engine it’s actually pretty terrible and very likely to distort an output that anybody who has been near a computer in the past thirty years can parse faster at a glance.
That is exactly the point, LLM aim to simulate the chaotic best guess flow of the human mind, to be conscious and at least present the appearance of thinking and from that to access and process facts but not be a repository of facts in themselves. The accusation here that the model constructed a fact and then built on it is missing the point, this is exactly the way organic minds work. Human memory is constantly reworked and altered based on fresh information and simple musings and the new memory taken as factual even while it is in large part fabricated, and to an increasing extent over time. Many of our memories of past events bear only cursory fidelity to the actual details of the events themselves to the point that they could be defined as imagined. We still take these imagined memories as real and act upon them exactly as has been done here by the AI model.
As below, stop with the analogies. No, that’s not “the chaotic best guess flow of a human mind”, that’s a whole bunch of tensor math generating likely chains of tokens. Those two things aren’t the same thing.
They aren’t the same thing in the strict sense, but they’re also not the same thing in practical terms at the end user level. If I ask a friend if they remember some half-forgotten factoid they can tell me not just if they do remember, but also how well they remember, how sure they are and why they know it. No LLM can do that, because LLMs know as little about themselves as about anything else. Which is nothing, because they’re LLMs, not people.