On 8/13/2026 9:44 AM, Andrew Pinski wrote:
On Thu, Aug 13, 2026, 7:13 AM Jeffrey Law
<[email protected]> wrote:
On 8/13/2026 3:38 AM, Richard Biener wrote:
> I think we're also seeing a lot of (unattributed) LLM assisted
patches lately.
Virtual certainty. The bits I've been using for reviews do
provide LLM
scoring for the main patch and the testsuite changes. I'm not using
that scoring for much at this time; I don't trust it yet and even
if it
did, the primary use would be to trigger me to look at the patch
deeper
on the assumption the LLM mucked up something.
>
> I have used LLMs to do offline bug analysis with moderate
success and
> I might want
> to play with LLMs taking up the daily churn of going through
bugzilla,
> adding missing
> classification, etc. - but that would make LLM modify bugs ....
Yea, I think bug classification, de-duplication and the like, etc
would
all be good uses. I'd bet that as the dust settles we could revisit
the guidelines with an eye towards this space.
I think bug reports should not be touched by llms. People are
expecting human responses.
That is the bugzilla changes should be added manually and not
interfaced from a llm. Support to users should be human interactions.
Otherwise it feels dry and folks will just ignore filing bugs.
I didn't really want to start hammering out that policy right now. But
it's a valid concern and I wasn't really envisioning LLMs interacting
with the reporter, but more for stuff that makes our jobs easier. Sure
the reporter could see that kind of activity (say adding the various
tags), but the primary interaction with the report should be human
(IMHO). But again this would be a space to work through in the next
iteration of LLM guidelines.
Jeff
