On Mon, Jul 27, 2026 at 11:12:05PM -0500, Matthias Geiger wrote: > If a human contributor writes a bad patch (by hand), I can point it out to > them, and tell them what they need to change. OTOH with LLM-generated > content I might as well flip a coin until something passable comes out. > Most people generating code like this can't even explain then what the code > does, which is not surprising. Furthermore, LLMs then to write overly long > code rather than concise, neat functions. > While even todays models with context can be mostly correct on the > code-aspect, they fail on (IMO) the most important aspect: social. LLMs can > never know, or feel, or (properly) gauge the social context, and tend to > censor topics that their authors thought to ban.
LLM-generator content is still sent by a human contributor. If you want to see an example of an LLM-generated patch, see [1]. As you can see in my patch review[2], I had a vague suspicion that it was LLM-generated, which the contributor acknowledged[3], and apologized for not including the Assisted-by tag. [1] https://lore.kernel.org/all/[email protected]/ [2] https://lore.kernel.org/all/[email protected]/ [3] https://lore.kernel.org/all/[email protected]/ Because Christian and I engaged the new contributor with respect, there is a possibility that Luyao might become a contributor member of the ext4 community. And indeed, that is one of the advantage of both AI-assisted tooling and AI-assisted review (which I did rely on in my response [2]) --- it can actually be a net positive in terms of engaging with potential new contributors. I've seen examples a new contributor was able to use the Sashiko review to learn how to improve their code, and the fact that the Sashiko review was able to give actionable suggestions on how to improve the code was something that the contributor reacted quite positively to. > Social interaction and context in code is a very important part in the free > software community and Debian, and LLMs do not know and do not care about > it. Emacs, vi, and Coccinelle scripts don't care about social interaction, either. But the humans who use those tools *do* care, and my experience is that using LLM as a tool can be helpful because it allows me to give more time and attention to these new contributors. As far as whether or not LLM's can understand code, in another one of my code reviews[4], I used an LLM to generate a 3000 word technical explanation about a particularly subtle aspect of ext4's block allocator[5]. I'm quite convinced that we don't need to worry about copyright concerns. Take a look at [5] and judge for yourself. [4] https://lore.kernel.org/all/[email protected]/ [5] https://lore.kernel.org/all/[email protected]/2-ext4_prealloc.md Was it perfect? No. But with some human editing, it will be useful code documentation that will help future human contributors --- and actually, future LLM code reviews and potentially future LLM coding agents that might assist in fixing bugs. And in the meantime it will be helpful to the new ext4 contributor who attempted to use Coccinelle script[6] which *clearly* didn't understand the context of the code that it was modifying. :-) [6] https://github.com/analogdevicesinc/linux/blob/main/scripts/coccinelle/api/atomic_as_refcounter.cocci Cheers, - Ted
