Because I was bored and wanted to see how close this thing was to taking my career from me, I decided to have Claude vibecode a kernel in Rust.
It did not suggest adding an allocator until it had gotten most of the way through creating kernel threads and context-switching.
Did I say "it did not suggest?" I had to tell it to because it was making statically-sized tables everywhere silently.
*golf clap*
Replying to @Elizafox@social.treehouse.systems
Also I told it to not let the kernel allocator panic in ordinary contexts but the allocator it made did. :D
Replying to @Elizafox@social.treehouse.systems
So how close is this thing to replacing my job?
Not a chance. lol
Replying to @Elizafox@social.treehouse.systems
Also unlike @ariadne's bunix which it made in a day because it was C and probably extremely well-trodden ground (could just steal shit from Linux + BSD), this has been a slog, comparatively speaking. It is like only 40% of the way to spawning anything in userspace for me. It took two days to even spawn a kernel thread, and it's still not done (returns aren't quite implemented).
Which admittedly is faster than I could have done it, but it has done this kind of incompetently and needs tons and tons of babysitting.
Replying to @Elizafox@social.treehouse.systems
I dunno. I don't get why companies want to mandate this crap on their workers when it produces obvious slop and can't do any kind of reasoning beyond basic applications, and to make LLMs scalable to the point of "reliably make a kernel" would take an astronomical amount of tokens and models with probably quadrillions of parameters. Which sounds fake, but no, that's about what it'd take.
Replying to @Elizafox@social.treehouse.systems
LLMs are worthless for software architectural work unless you're extremely dilligent and know where to poke. No manager is going to replace a senior dev with this. No manager would know what to ask, what to say, what to push back on, just silently say yes. The thing is LLMs can get 90% of things right, it's that last 10% that matters, and that 10% is not "obscure things a normal programmer would miss", it's "very obvious glaring holes it just ignored."
