Submitted 11 hours ago by Lemmynated@lemmy.zip to technology@lemmy.zip
I knew this would happen, I talked about it with everyone for a year. LLMs need to be used more carefully and cannot replace workers, unless you want a far more expensive worker.
This is why I think Mistral AIs focus on small domain specific models that will start paying off once companies realize that renting a really expensive general model to do specific work is a waste of money compared to a small, possibly local and owned, model can do at a fraction of the cost.
Exactly, the future is a bunch of smaller task specific models that can be called rather than one massive general use model. It should save a lot of power in terms of inference as well.
The real world use-cases are not what techbro CEOs promised? WHO would have thought?
The ”quality” of current gen LLMs limits their usage to secondary tasks like pure code generation, documentation or testing. It’s impactful but nowhere near the point where you start replacing your engineers with it.
Harness it for very specific tasks and maybe you can gain productivity increases but for anything else its going to cost a lot and still need constant handholding
stoy@lemmy.zip 11 hours ago
The one useful feature of AI that I have found is to use it as a dynamic manual.
I ask about specific commands, or how to do something, similar to a normal search engine.
I then test the answer given, and if successful, I implement the knowledge into my code.
Something like this:
How do I make a progress bar with diagonal stripes in HTML/CSS?
I get a generic answer and start experimenting with it.
I could easily skip the AI part and go back to regular searching, it would slow me down, but I would be fine + it would be less wasteful.
WhatAmLemmy@lemmy.world 7 hours ago
It’s honestly amazing if used correctly. Having to remember awk, sed, etc, the exact syntax, idiosyncrasies across OS’s, was always a pain the ass for me. It often makes mistakes, but less than I do trying to customize solutions from the internet. I get a better result much quicker.
My concerns with unsupervised learning statistical models have never been about the tech. It’s always been about capitalism and how the psychopaths in charge would wield it.