On Tue, Oct 27, 2026 at 10:33:24AM -0500, Ben Finney wrote:
> Martin Steigerwald <[email protected]> writes:
>
> > One of the greater challenges with LLMs seems to be to guide them to
> > produce concise, accurate and to the point output.
>
> Until that day comes (if ever), I agree with other posters: This
> unreliable and non-deterministic machine should not be sending messages
> unsolicited to Debian package maintainers.
Humans are unrelible and non-deterministic, you know. The real
question is whether advisory messages are useful enough if they will
be helpful to developers / engineers / maintainers.
In the kernel world, the vast majority of subsystem maintainers have
elected to enable AI generated code reviews. They have the option of
only having the reviews go to a website, or just to the maintainer, or
to the maintainer and patch submitter, or to the mailing list. Over
time, most kernel subsystem maintainers have decided despite the fact
that the reviews are not perfect; sometimes there are false positives;
sometimes there are real issues which are missed; sometimes there are
issues that didn't get picked up in the v1 patch (even though the
problem was there), but only got picked up in the review of the v2
patch series. Yes, it's all annoying --- but these faults its exist
with human reviewers as well, and most maintainers, after looking at
the result, have decided that the benefits vastly outweigh the
downsides of false positives reports.
And yes, people can ignore the AI generated reports. But the
automated reviews may find problems before the human reviewers have
time to do their code review. And if the submitters can fix the real
issues which are pointed out by the AI review, they can quickly come
up with a revised submission, and then in the end it saves everyone
time.
So if you think it's a waste of time, feel free to ignore the report.
But you might find that if it's pointing out things that humans later
have validated to be true, wouldn't you rather have first access if it
allows your NEW package to get accepted faster?
- Ted
