Re: [BUG] hw/xen: features published after InitWait since 240cc11369fc

[David Woodhouse <dwmw2@xxxxxxxxxxxxx>]

On Fri, 2026-09-11 at 15:56 +0100, Peter Maydell wrote:
> On Fri, 11 Sept 2026 at 15:39, David Woodhouse <dwmw2@xxxxxxxxxxxxx> wrote:
> > I wasn't *actually* setting out to comment on the AI policy today  — in
> > fact I didn't really have a strong opinion on it. I spend enough of my
> > life tilting at windmills *intentionally*; this wasn't meant to be one
> > of them.
> > 
> > Letting the tool reply in its own voice and pretending I was held
> > hostage was mostly meant as a joke, as *well* as being the only way I
> > was going to look at that bug today (it would have taken me a *long*
> > time to do all that testing of old and new failure modes, finding the
> > net device path that *does* reproduce it locally, fixing the other
> > problem with that, and finally confirming that it all works).
> 
> Even our current strict AI policy is fine with using them as tools
> to help in debugging, testing, and so on. What I am complaining
> about is not that you did that, but that you posted an enormous
> long apparently unedited pile of output from the LLM, rather than
> using it to do the work and then posting your conclusions. That
> is asking readers of the thread to do a lot of work for you
> (i.e. wading through the huge email to figure out whether the LLM
> output is right, wrong or irrelevant).

What makes you think I hadn't already iterated through it and ensured
that was it was doing was right, relevant *and* sufficient? Did I miss
something?

(On this occasion, its original attempts were mostly lacking in the
latter; I had to explicitly make it go and check on the things which
were called out as open questions in the original mail. I cancelled its
first draft after reading it, and made it try again.)

Note that saying things which are wrong, irrelevant and insufficient is
not solely the domain of AI. We see plenty of that from real people
too. Bug reports, especially, have been the target of jokes for
*decades* already. My favourite is still "dirty like zebra"¹.

I also *did* preface that email with my conclusion agreeing that the
solution was indeed to move one line of code, as well as an explicit
statement that the remainder *was* the direct output (after numerous
iterations of guidance) of the tool, covering all the details and open
questions from the original.

That original *also* looks very much like AI was involved in its
drafting as well as the investigation, FWIW. Which is partly why there
was so much to reply *to*.

And let's look again at what my message *did* say:

| I've done the gruntwork of confirming your diagnosis, re-testing the
| 2023 crash that motivated commit 240cc11369fc, and verifying that the
| fix shape you describe (defer only the InitWait transition until after
| the implementation's realize method) addresses both without
| reintroducing either.

Even if someone didn't stop reading at "Claude here", that is actually
fairly succinct given that I deliberately *didn't* rein it in or
rewrite it this time, or even tell it "use a tenth of the words" as I
often do. And then the reader has another chance to stop reading before
it goes on with its "Summary of what was established" which confirms
for the record that it *did* actually do its due diligence.

I deliberately *didn't* reword that message, and I *did* clearly demark
it for what it was, but honestly I wouldn't be embarrassed to have
posted that as my own wording.

It's just a tool. It's as useful as the person controlling it.

> > But our policies should be based on actual data and the holistic
> > outcomes they achieve, and this is just data which we can take into
> > account on one side of the balance, even if the other side remains more
> > compelling.
> 
> Yes. I view that email as pretty strong data for 'even a policy
> shift which says "we're OK with accepting some kinds of AI
> generated code" should still be pretty strong on "text intended
> for humans to read (docs, commit messages, mailing list posts, etc)
> should be written by humans"'. I think that review commentary on
> Paolo's RFC about a less strict policy tended to be in that direction,
> so I don't think I'm completely out on a limb here.

Great. Happy to help provide real world data.

Perhaps a variant to consider might be "text generated by AI must be
clearly labelled as such" rather than barring it outright?



¹ https://bugzilla.redhat.com/show_bug.cgi?id=61350

Attachment: smime.p7s
Description: S/MIME cryptographic signature