Linux

They're getting better fast

https://www.xda-developers.com/qwen-3-8-27b-reverse-engineering-job-frontier-model/

Still unreliable for tasks which aren't easy to validate in code, though

I mean the best local models are only a few months behind the the best proprietary cloud models. Just look at Qwen 3.8 Flash or 27B. Imo local models are more than sufficient if you're going to use AI for programming. Who cares if the small model can't oneshot the whole patch, you shouldn't be submitting raw LLM code to public repos anyway. Imo the only acceptable use of LLMs for coding is as a way of rapidly prototyping ideas that you will later mostly/entirely rewrite by hand, or as an extra static analysis tool for finding potential security holes.

And as someone who frankly doesn't give a shit about intellectual property over code, my main ethical issue with LLMs is the monstrous resource consumption of hyperscale datacenters, so local models are strongly preferable. Also for privacy reasons.

It looks like their point is that they don't want Debian's codebase to be used to train corporate AI models, and almost all the proposals seem to agree on that at the very least. I feel like a required AI disclosure would have been better, but what do I know, I'm not a Debian contributor

You are not supposed to use online LLMs for undisclosed security stuff; there is no general ban of online services if you can take responsibility for the stuff it produces for the contributor. At least, this is how I read the general resolution.

This is one more reason, that using Debian as my main distro was a good choice.