There are some very useful use cases. One that I have a decent amount of experience in is upscaling SD (or even VHS) content. Depending on the quality of the source material, you can get very good results.
I’ve also found LLMs helpful as a compliment to web searching for my work (I deal with lots of public datasets from international organisations); LLM queries have helped me find sources that I missed via directed search.
It can also be helpful as guided learning/reference system for Linux CLI (I tend to forget the parts I rarely use) or even software application more broadly (used it to help learn about GIS applications that I needed to use to access historical weather data for a work project).
So “as a compliment to web search” is a great example. We had really good we search based off deterministic algorithms. Now we have enshittified web search, so you boost it with AI that was significantly more expensive to train, significantly more expensive to run inference on, and can confidently return false “facts”.
We’re not using it because it’s better. We’re using it because the alternative got worse.
I’ve also found LLMs helpful as a compliment to web searching for my work (I deal with lots of public datasets from international organisations); LLM queries have helped me find sources that I missed via directed search.
Do consider that web search had been a computationally simple and more importantly solved problem that has gotten steadily more unsolved over the course of the past 25 years.
Google web search (and by extension services like Ecosia that rely on Google) has worsened in quality, especially for more niche content that requires some work to get to.
They’re novelties for sure, but I see nothing but toy uses or things that be achieved by other, less wasteful, techniques.
I assume by crypto you mean cryptocurrencies rather than any cryptography, because there’s lots of valid uses for private communication.
Yes, I didn’t mean cryptography of course. :)
There are some very useful use cases. One that I have a decent amount of experience in is upscaling SD (or even VHS) content. Depending on the quality of the source material, you can get very good results.
I’ve also found LLMs helpful as a compliment to web searching for my work (I deal with lots of public datasets from international organisations); LLM queries have helped me find sources that I missed via directed search.
It can also be helpful as guided learning/reference system for Linux CLI (I tend to forget the parts I rarely use) or even software application more broadly (used it to help learn about GIS applications that I needed to use to access historical weather data for a work project).
So “as a compliment to web search” is a great example. We had really good we search based off deterministic algorithms. Now we have enshittified web search, so you boost it with AI that was significantly more expensive to train, significantly more expensive to run inference on, and can confidently return false “facts”.
We’re not using it because it’s better. We’re using it because the alternative got worse.
Do consider that web search had been a computationally simple and more importantly solved problem that has gotten steadily more unsolved over the course of the past 25 years.
There is a lot of truth to that.
Google web search (and by extension services like Ecosia that rely on Google) has worsened in quality, especially for more niche content that requires some work to get to.