@nprofile1q... Ugh. This is why I'm a fillename purist. No spaces, no special characters. Just... ugh always eventually causes problems somewhere, somehow.

@nprofile1q... To be clear, I mean it was like mentoring HS students in that they make simple, but reasonable mistakes constantly. With students I can usually guide them towards understanding.

With LLMs, they just confidently make the same mistakes over and over again. When they finally "understand" what their error was they get all "Oh, you're right!" It's *maddening*.

I've tried several times to be "open minded" about it, and I just can't. I come away so frustrated every time.

@nprofile1q... To be clear, I mean it was like mentoring HS students in that they make simple, but reasonable mistakes constantly. When you try and get them to understand them, sometimes they just have no idea what you are talking about.

Like I'll explain or remind them of how things work, enough to let them try to work out the rest. When I do that with LLMs they just... ugh. It doesn't work. They chatter like they are smart, but they make the dumbest mistakes over and over. I don't have the patience.

@nprofile1q... I will say some of the bigger local models like the 31b param Gemma 4 fun as toys for interactive stories. The fact that they *always* create the same characters, and use the same adjectives over and over tells me a lot though. (even between completely different models)

It's bizarre that they seem to output fairly complicated emotional reactions, work with subtle metaphors, and yet completely forget where you are, or who knows what. I'm 100% convinced vibe coding is the same.

@nprofile1q... Each of them literally wrote half a novel (100k+ words) of "thinking", and still just could never do any of the simple problems I tried to make them fix.

Honestly, I went in to it with a dismissive attitude, so maybe that was my "problem". 🤷 The whole thing felt *exactly* like mentoring high school students to program, which I've done for several years now. The one even argued with me that you could use a variable before you defined it because forward references existed. >_<

@nprofile1q... Local models are actually surprisingly good at understanding/explaining code (eh... kinda?), but I tried having them fix bugs that I already fixed and they just... couldn't. In the one case I tried a half dozen 30+ billion param models to implement a simple feature in my Lua debugger. Something I did in 2 lines of code in 5 minutes. I gave them all *hours* to do it and they just... couldn't. The closest I got was that one hand inlined a function that was ~5% of the total debugger size.

@nprofile1q... Uff. I have complex feelings about this. On the one hand local LLMs being a free public good that benefits everyone equally fixes many of my problems with how the data is basically stolen from "regular" people and then sold back to us. (IP rights are *not* for us!) On the other hand, I've tried them and found them to be basically useless. I'm in the rare minority of people that have access to enough RAM to run models that need 80+ GB of RAM and... uff they are still kinda just useless.

@nprofile1q... I wanted to have a witty reply to this but all I could think of was (fshhhh… crackle)

@nprofile1q... I’m glad that something in your life sucks right now. You deserve it. 🤔

@nprofile1q... In 2026 it feels ironic to need optical media for a wireless adapter. >_<