I think a better approach is needed than a complete ban on AI contributions.

Just throwing such contributions out is a big loss.

I have gained a lot of experience working for over 20 years in our industry, but writing CUDA Kernels in not one of my skills or something I'd want to spend weeks to learn.

So recently when I was trying to run a particular local AI model and the inference software was running pretty badly I haven't even considered debugging it myself. It would take me weeks to even start properly.

But it was Saturday evening and my Claude Code usage resets on Sunday and I haven't used any of my Fable 5.1 quota yet this week. In fact I found no need to use fable at all, opus doing everything I needed.

So I decided to let fable work on a problem overnight. Llama.cpp, fix the bugs that cause horrible speed decay in Qwen3.8-Flash-Next in longer context and with multiple users. I told it it can build llama, run it on my hardware and iterate as much as possible.

In the morning I found it finished. It managed to produce up to 10x speedup in some cases (running 5 simultaneous users, with long contexts). 6 thousand lines of code. 19 patches. A ridiculous amount of work. Each patch explained in a readme. Speed tests run before and after for each.

This code made a huge difference for me. Suddenly I have a local model almost as capable as opus that I can run in 5 simultaneous claude code/pi sessions with further 5 sessions idle and cached in RAM.

I've run that code for last 3 weeks. It hasn't crashed once. It hasn't failed me once. Based on it I created a system where fable, limited to 200k of contexts only distributes work to 5 Qwen agents. This way I could do far more of such type of work (including fixing vllm and adding missing features to it in a similar way).

But would I ever post my 19 patches as a PR to llama.cpp? Hell no. I do not know what other models they will break. How they will behave on other people's hardware. Or how contrary to architecture authors have in mind they are.

So what did I do? I want those fixes to make their way to llama.cpp eventually.

I published it as a "temporary fork". I posted about it on the llama.cpp discussions. And I hope authors will use that code if they want to speed that model up.

If not, that is fine too. People like me that can benefit from those fixes also have access. And I'm not risking my professional reputation by submitting something I haven't read and fully understood.

Stop this typical verbose damage control. This is Jehovah's witness level or worse.

Come to Jesus, come to AI, written in an engaging novelist style. Maybe sell the rights to Hollywood.

Reality: No software has gotten better since AI, and only the tinkerers and learning-resistant cherish AI and promote it like their favorite computer game.