LLMs shouldn’t be used to search for what’s right or wrong, they should be used to quickly get a wide breadth of information on a given topic or provide a starting point on code or text projects that you can refine later.
For example, I wanted to use a new library in a project, and the documentation was a bit cryptic and examples weren’t quite right either. So I asked an LLM tuned for coding tasks to generate some example code for our use case (described the features I wanted to use), and the code worked. I needed to tweak some stuff (like I would w/ any example), but it worked. I used the LLM because I knew there would be a bunch of public code projects using this library with a variety of use cases, and that the LLM would probably do a reasonable job of picking out a decent one to build from, and I was right.
On another topic, I needed to do research on an unfamiliar topic, so I asked the LLM to provide a few examples in that domain w/ brief descriptions. That generated a ton of useful keywords that I used to find more reputable sources (keywords that would’ve taken hours of searching to generate), so I was able to quickly find reliable information starting from a pretty vague notion of what I wanted to learn.
LLMs have a lot of limitations, but if they’re used to accomplish common tasks quickly, they can be incredibly useful. I don’t think they’ll replace anyone’s job (unless your job is already pointless), traditional search engines (as much as Google wants it to), or anything like that, but they are useful tools that can make some of the more annoying parts of my job more efficient.
LLMs have flat out made up functions that don’t exist when I’ve used them for coding help. Was not useful, did not point me in a good direction, and wasted my time.
Sure, they certainly can hallucinate things. But some models are way better than others at a given task, so it’s important to find a good fit and to learn to use the tool effectively.
We have three different models at work, and they work a lot differently and are good at different things.
Then you’re using it wrong.
LLMs shouldn’t be used to search for what’s right or wrong, they should be used to quickly get a wide breadth of information on a given topic or provide a starting point on code or text projects that you can refine later.
For example, I wanted to use a new library in a project, and the documentation was a bit cryptic and examples weren’t quite right either. So I asked an LLM tuned for coding tasks to generate some example code for our use case (described the features I wanted to use), and the code worked. I needed to tweak some stuff (like I would w/ any example), but it worked. I used the LLM because I knew there would be a bunch of public code projects using this library with a variety of use cases, and that the LLM would probably do a reasonable job of picking out a decent one to build from, and I was right.
On another topic, I needed to do research on an unfamiliar topic, so I asked the LLM to provide a few examples in that domain w/ brief descriptions. That generated a ton of useful keywords that I used to find more reputable sources (keywords that would’ve taken hours of searching to generate), so I was able to quickly find reliable information starting from a pretty vague notion of what I wanted to learn.
LLMs have a lot of limitations, but if they’re used to accomplish common tasks quickly, they can be incredibly useful. I don’t think they’ll replace anyone’s job (unless your job is already pointless), traditional search engines (as much as Google wants it to), or anything like that, but they are useful tools that can make some of the more annoying parts of my job more efficient.
LLMs have flat out made up functions that don’t exist when I’ve used them for coding help. Was not useful, did not point me in a good direction, and wasted my time.
You need to actively have the relevant code in context.
I use it to describe code from shitty undocumented libraries, and my local models can explain the code well enough in lieu of actual documentation.
Sure, they certainly can hallucinate things. But some models are way better than others at a given task, so it’s important to find a good fit and to learn to use the tool effectively.
We have three different models at work, and they work a lot differently and are good at different things.