A few days ago I wanted to change something on my local media server.
Normally I'd SSH into it, remember where I'd configured everything, look up a couple of commands I'd forgotten, make the change and move on.
Instead I opened Pi(coding agent) and typed, more or less: Find out why this isn't working and fix it.
Behind Pi was Qwen3.6-35B-A3B-Q4_K_M running on my own machine.
A few minutes later the job was done.
There is something about this that still feels absurd to me.
I grew up watching Dexter's Laboratory. Dexter would sit inside his secret lab, talk to computers and machines, and somehow the whole laboratory would respond to him.
Obviously my Linux box isn't Dexter's lab. But the interaction is becoming strangely familiar.
I tell my computer what I want rather than precisely how to do it.
I've been experimenting with local models for quite a while now, but only recently they have become useful on the kind of hardware I have. I'm not handing one my serious codebases and telling it to autonomously build features. The models aren't reliable enough for me to trust them that way.
But I've found another category where they're becoming extremely useful:
things I know I could do myself, but don't particularly want to do.
A Bash command. A configuration change. Investigating some service on my media server. A small script. Understanding some obscure error. All this while keeping things private.
The model doesn't need to be a genius for these jobs. It needs to be competent enough that supervising it is easier than doing everything myself.
Another use case is for language learning - the spoken ones - I keep a window to Local LLM open, so that I can ask it meaning of certain words, give me translations, help me refine my language etc, without ever sending my text across the fibre.
And I think that's the part I'm finding more interesting than the models themselves.
For most of my life, operating a computer meant learning its language. Commands. Menus. Configuration files. APIs. We're slowly adding another layer where the computer learns enough of our language that we can meet somewhere in the middle. Today I have a reasonably capable model running locally, operating my own computer under my supervision. And most of the time I use this little piece of science fiction because I'm too lazy to remember a Bash command.