I've run into major problems with LLMs as I maintain my home Linux systems. If I just copied commands they list, I would have a near 100% failure rate, as most of the information they have has been gleaned from forum posts that are years out of date.
They're a good jumping off point, but I need to delve into the original sources just like I did when I used Google.
I'm not sure. I run a homelab tailscale/k3s setup with gitops, dozens of services, VMs, backups, etc, all vibed by claude, and works just fine. Didn't write a single line of code for this. I don't know kubernetes and never will.
> I don't know kubernetes and never will.
He said, proud of his own ignorance.
Proud? No, but I am not ashamed of it also.
I would be the first to admit my ignorance on the absolute majority of topics. There is a limited number of things I can learn in life, and kubernetes won't be one of them - I'm just not interested in it (and all the other infra stuff, to be honest), as long as it works.
What on earth LLM are you using and do you tell it which distro/version of Linux they are supposed to be working with + give them access to web search or man pages? I haven't had this issue in like a year and a half.
By default I use Leo, and I do specify the distro. I tried to qualify by version, but when I do that I lose out on a lot of correct information for things that haven't changed in awhile.
I've had the opposite experience giving claude code SSH/ADB access to my devices. Not sure if it's the model itself or the harness, but I haven't had to manually do a sysadmin task in months.
Have a look at https://github.com/ThorOdinson246/whatisit-nl2sh
It seems to get shell scripts right most of the time.