None of what you brought up as a positive are things an LLM does. Most of those things existed before the modern transformer-based LLMs were even a thing.
LLM-s are glorified text prediction engines and nothing about their nature makes them excel at formal languages. It doesn’t know any rules. It doesn’t have any internal logic. For example if the training data consistently exhibits the same flawed piece of code then an LLM will spit out the same flawed piece of code, because that’s the most likely continuation of its current “train of thought”. You would have to fine-tune the model around all those flaws and then hope some combination of a prompt won’t lead the model back into that flawed data.
I’ve used LLMs to generate SQL, which according to you is something they should excel at, and I’ve had to fix literal syntax errors that would prevent the statement from executing. A regular SQL linter would instantly pick up that the SQL is wrong but an LLM can’t pick up those errors because an LLM does not understand the syntax.
XLE@piefed.social 1 day ago
Citation needed.
You’re on a post about Linux, an OS has grown in popularity thanks to Microsoft ruining Windows with the “true aids” you’re promoting here.
dukemirage@lemmy.world 1 day ago
Whatever MS bakes into Windows is not what I listed above. Spin up a local LLM trained on your code base and try using it.
XLE@piefed.social 1 day ago
No thanks AI bro
dukemirage@lemmy.world 1 day ago
Well I will not share a screencast where a local LLM helps with code completion on a private project. You talk like you’re a proficient developer, you can try that on your own. And where is the fallacy?
tjsauce@lemmy.world 1 day ago
Hey I’m against corporate AI too, but when anyone can create a very basic ML program that runs locally with public domain data, eventually something both useful and ethical will emerge. It’s good to be skeptical, but you don’t have to be an AI bro to see that some specific tools might meet or exceed your standards.
I don’t like image or video generators, but the core tech is really useful for frame interpolation, a usecase that is not inherently controversial and badly needs improvement.
Sorry to not-x-it’s-y, but it’s not about forcing the big tool into your workflow, it’s about finding the 1001 little tools that work every time and collecting them. Or, wait for these tools to be consolidated.
If I seem naive, It’s cause I believe in reclaiming as much from tainted technology as possible.