I use LLMs before search especially when I’m exploring all possibilities, it usually gives me some good leads.
I somehow know when it’s going to be accurate or when it’s going to lie to me and I lean on tools for calculations, being time aware, and web search to help with the lies.
Sure but you can benchmark accuracy and LLMs are trained on different sets of data using different methods to improve accuracy. This isn’t something you can’t know, and I’m not claiming to know how, I’m saying that with exposure I have gained intuition, and as a result have learned to prompt better.
Ask an LLM to write powershell vs python, it will be more accurate with python. I have learned this through exposure. I’ve used many many LLMs, most are tuned to code.
Currently enjoying llama3.3:70b by the way, you should check it out if you haven’t.
The same can be said about the search results. For search results, you have to use your brain to determine what is correct and what is not. Now imagine for a moment if you were to use those same brain cells to determine if the AI needs a check.
AI is just another way to process the search results, that happens to give you the correct answer up front, most of the time. If you go blindly trust it, that’s on you.
If you knew what the sources were, you wouldn’t have needed to search in the first place. Just because it’s on a reputable website does not make it legit. You still have to reason.
You are spending more time and effort doing that than you would googling old fashioned way. And if you don’t check, you might as well throwing magic 8-ball, less damage to the environment, same accuracy
No it doesn’t. It incorporates unknown number of words from the internet into a machine which only purpose is to sound like a human. It’s an insanely complicated machine, but the truthfulness of the response not only never considered, but also is impossible to take as a deaired result.
And the fact that so many people aren’t equipped to recognise it behind the way it talks could be buffling, but also very consistent with other choices humanity takes regularly.
When it’s important you can have an LLM query a search engine and read/summarize the top n results. It’s actually pretty good, it’ll give direct quotes, citations, etc.
And some of those citations and quotes will be completely false and randomly generated, but they will sound very believable, so you don’t know truth from random fiction until you check every single one of them. At which point you should ask yourself why did you add unneccessary step of burning small portion of the rainforest to ask random word generator for stuff, when you could not do that and look for sources directly, saving that much time and energy
I guess it depends on your models and tool chain. I don’t have this issue but I have seen it for sure, in the past with smaller models no tools and legal code.
As a side note, I feel like this take is intellectually lazy. A knife cannot be used or handled like a spoon because it’s not a spoon. That doesn’t mean the knife is bad, in fact knives are very good, but they do require more attention and care. LLMs are great at cutting through noise to get you closer to what is contextually relevant, but it’s not a search engine so, like with a knife, you have to be keenly aware of the sharp end when you use it.
I ask GPT for random junk all the time. If it’s important, I’ll double-check the results. I take any response with a grain of salt, though.
So, if it isn’t important, you just want an answer, and you don’t care whether it’s correct or not?
I use LLMs before search especially when I’m exploring all possibilities, it usually gives me some good leads.
I somehow know when it’s going to be accurate or when it’s going to lie to me and I lean on tools for calculations, being time aware, and web search to help with the lies.
Are you familiar with Dunning-Kruger?
Sure but you can benchmark accuracy and LLMs are trained on different sets of data using different methods to improve accuracy. This isn’t something you can’t know, and I’m not claiming to know how, I’m saying that with exposure I have gained intuition, and as a result have learned to prompt better.
Ask an LLM to write powershell vs python, it will be more accurate with python. I have learned this through exposure. I’ve used many many LLMs, most are tuned to code.
Currently enjoying llama3.3:70b by the way, you should check it out if you haven’t.
The same can be said about the search results. For search results, you have to use your brain to determine what is correct and what is not. Now imagine for a moment if you were to use those same brain cells to determine if the AI needs a check.
AI is just another way to process the search results, that happens to give you the correct answer up front, most of the time. If you go blindly trust it, that’s on you.
With the search results, you know what the sources are. With AI, you don’t.
If you knew what the sources were, you wouldn’t have needed to search in the first place. Just because it’s on a reputable website does not make it legit. You still have to reason.
You are spending more time and effort doing that than you would googling old fashioned way. And if you don’t check, you might as well throwing magic 8-ball, less damage to the environment, same accuracy
The latest GPT does search the internet to generate a response, so it’s currently a middleman to a search engine.
No it doesn’t. It incorporates unknown number of words from the internet into a machine which only purpose is to sound like a human. It’s an insanely complicated machine, but the truthfulness of the response not only never considered, but also is impossible to take as a deaired result.
And the fact that so many people aren’t equipped to recognise it behind the way it talks could be buffling, but also very consistent with other choices humanity takes regularly.
False.
When it’s important you can have an LLM query a search engine and read/summarize the top n results. It’s actually pretty good, it’ll give direct quotes, citations, etc.
And some of those citations and quotes will be completely false and randomly generated, but they will sound very believable, so you don’t know truth from random fiction until you check every single one of them. At which point you should ask yourself why did you add unneccessary step of burning small portion of the rainforest to ask random word generator for stuff, when you could not do that and look for sources directly, saving that much time and energy
I, too, get the feeling, that the RoI is not there with LLM. Being able to include “site:” or “ext:” are more efficient.
I just made another test: Kaba, just googling kaba gets you a german wiki article, explaining it means KAkao + BAnana
chatgpt: It is the combination of the first syllables of KAkao and BEutel - Beutel is bag in german.
It just made up the important part. On top of chatgpt says Kaba is a famous product in many countries, I am sure it is not.
I guess it depends on your models and tool chain. I don’t have this issue but I have seen it for sure, in the past with smaller models no tools and legal code.
As a side note, I feel like this take is intellectually lazy. A knife cannot be used or handled like a spoon because it’s not a spoon. That doesn’t mean the knife is bad, in fact knives are very good, but they do require more attention and care. LLMs are great at cutting through noise to get you closer to what is contextually relevant, but it’s not a search engine so, like with a knife, you have to be keenly aware of the sharp end when you use it.