AI can be a useful tool and I think it will slowly become more common in the workplace, for example it can be very convenient for knowledge retrieval, but it’s laughable to think that it can replace humans. I’d wager any time “AI” can replace a human the job could’ve already been automated through other means.
Generalized LLMs like ChatGPT are. If you train a model on your own documentation then all it “knows” is what is in the docs and it can perform very well at finding relevant results. It’s just kind of a context-aware search engine at that point.
The problem again is that companies mostly aren’t doing that, they’re trying to replace humans with ChatGPT.
Perhaps I’m using the wrong terminology. But being able to ask in natural language “why is something the way it is” and it returns references to code, bugs, and documentation along with a small summary is pretty cool. It works better than any of the half-baked corporate search engines I’ve used before. Is this not “knowledge retrieval”? In any case I can see the utility.
Sure, you can’t trust LLMs and just copy-paste whatever comes out of it. But it’s very effective as a way to find something in very large mixed datasets when you may not know which exact keywords to use for a traditional search engine.
Yeah the AI hype levels are insane, but at the same time I think there is some interesting and actually useful technology there. That’s my 2c anyway.
The search thing is specific to internal data sets btw. Anyone who has used intranet search engines at large companies would probably relate just how terrible they are. Much worse than Google is at searching the internet.
Anyone who has used intranet search engines at large companies
Sharepoint search functionality comes to mind. Our team commonly refers it as write-once storage as once you throw something in there you’ll never find it again. And yes, we stole the term from somewhere.
“Fixes companies internal documentation” is actually a huge get for AI, and would be worth some real hype, but yeah.
That’s still peanuts compared to the marketing, which is why people are getting pretty tired of the whole AI push. The actual, incremental improvements are being run over roughshod by snake oil salesmen.
I lived through the dotcom hype cycle, the 5G cycle, the crypto cycle, etc. The useful (boring!) bits of technology remain and something new and shiny becomes the target of hype and speculation a few years later. Nothing new really.
Eh. It’s useful for finding what I want to know. The result to a query which goes like “Based on this paragraph from some documentation written in 2005 (link) the answer is <bunch of generated text rehashing the information I wanted to find in the first place>” is a whole lot more useful than “Here is a list of thousands and thousands of irrelevant and incoherently sorted results, of which one is probably what you were looking for. Good luck.” which was, unfortunately, the state of the art up to this point.
AI can be a useful tool and I think it will slowly become more common in the workplace, for example it can be very convenient for knowledge retrieval, but it’s laughable to think that it can replace humans. I’d wager any time “AI” can replace a human the job could’ve already been automated through other means.
LLMs are absolute garbage for knowledge retrieval.
Generalized LLMs like ChatGPT are. If you train a model on your own documentation then all it “knows” is what is in the docs and it can perform very well at finding relevant results. It’s just kind of a context-aware search engine at that point.
The problem again is that companies mostly aren’t doing that, they’re trying to replace humans with ChatGPT.
Except that your context aware search engine would tell you when there is no result and AI will just make shit up and distort the results it did find.
It’s not true.
Vector dbs and LLMs are really powerful at knowledge retrieval.
See notebooklm and open-source alternative.
They are “media transformers” and might be useful if limited to it.
Knowledge retrieval certainly not, as “they” know nothing besides how likely one data fragment is to appear near other data fragments.
Perhaps I’m using the wrong terminology. But being able to ask in natural language “why is something the way it is” and it returns references to code, bugs, and documentation along with a small summary is pretty cool. It works better than any of the half-baked corporate search engines I’ve used before. Is this not “knowledge retrieval”? In any case I can see the utility.
I’ve tested it with both python and Cisco iOS pretty thoroughly and it very convincingly gets things wrong a lot.
The problem with that is the constant hallucinations and complete lack of correctness checks.
Sure, you can’t trust LLMs and just copy-paste whatever comes out of it. But it’s very effective as a way to find something in very large mixed datasets when you may not know which exact keywords to use for a traditional search engine.
So your overall point is that AI is a better search engine. “It’s like google, but better.”
This is both likely true, and no where near fantastical enough to justify the trillion dollar hype cycle.
Yeah the AI hype levels are insane, but at the same time I think there is some interesting and actually useful technology there. That’s my 2c anyway.
The search thing is specific to internal data sets btw. Anyone who has used intranet search engines at large companies would probably relate just how terrible they are. Much worse than Google is at searching the internet.
Sharepoint search functionality comes to mind. Our team commonly refers it as write-once storage as once you throw something in there you’ll never find it again. And yes, we stole the term from somewhere.
“Fixes companies internal documentation” is actually a huge get for AI, and would be worth some real hype, but yeah.
That’s still peanuts compared to the marketing, which is why people are getting pretty tired of the whole AI push. The actual, incremental improvements are being run over roughshod by snake oil salesmen.
I lived through the dotcom hype cycle, the 5G cycle, the crypto cycle, etc. The useful (boring!) bits of technology remain and something new and shiny becomes the target of hype and speculation a few years later. Nothing new really.
So at best it might be useful for hinting.
Eh. It’s useful for finding what I want to know. The result to a query which goes like “Based on this paragraph from some documentation written in 2005 (link) the answer is <bunch of generated text rehashing the information I wanted to find in the first place>” is a whole lot more useful than “Here is a list of thousands and thousands of irrelevant and incoherently sorted results, of which one is probably what you were looking for. Good luck.” which was, unfortunately, the state of the art up to this point.
When I try that, I just get confidently incorrect answers.
my dad said the same about excel. it just gives him incorrect results compared to his handwritten sheet. now let’s get you to bed grandpa.