Conversation
Jarkko Sakkinen
jarkko
I've never even really considered using AI for writing kernel patches.
We could go through the usual reasons, like simply congitive abilities to be a gatekeeper for anything but there are also cost-benefit measures that simply make using AI a bad investment in this particular area of expertise.
Writing the code takes about 1-5% of the all time spent on a kernel patch. It is still really important meditation time too, as then you can go through your plan/strategy slowly and make sure every brick is in the right place. I notice tons issues while writing code even if the overall plan is solid as **** (... and this is why BTW spec-driven agent loop development is complete and uttermost nonsense).
By cutting that 1-5% there is at least no financial at all, and the drawback is a constant ffux of kernel bugs because of *mistakes*, not unexpected conditions. Even if it was 0,1% that a frontier LLM generates a kernel bug, it is 0,1% too much. And even if I can human-check and test it (which I would, it is not at all the same than the quality time spent browsing the kernel tree :-)
We could go through the usual reasons, like simply congitive abilities to be a gatekeeper for anything but there are also cost-benefit measures that simply make using AI a bad investment in this particular area of expertise.
Writing the code takes about 1-5% of the all time spent on a kernel patch. It is still really important meditation time too, as then you can go through your plan/strategy slowly and make sure every brick is in the right place. I notice tons issues while writing code even if the overall plan is solid as **** (... and this is why BTW spec-driven agent loop development is complete and uttermost nonsense).
By cutting that 1-5% there is at least no financial at all, and the drawback is a constant ffux of kernel bugs because of *mistakes*, not unexpected conditions. Even if it was 0,1% that a frontier LLM generates a kernel bug, it is 0,1% too much. And even if I can human-check and test it (which I would, it is not at all the same than the quality time spent browsing the kernel tree :-)
1
1
4
... and I like browsing kernel, I do enjoy it too...
When it comes to code, especially C code, there really isn't itch with it that I've wanted to get rid of years or decades... Dev tools are better than ever and bpftrace especially is awesome. It's all fun.
When it comes to code, especially C code, there really isn't itch with it that I've wanted to get rid of years or decades... Dev tools are better than ever and bpftrace especially is awesome. It's all fun.
0
0
3