On Fri, 30 Jan 2026 at 20:28, Chris Smith <[email protected]> wrote:
>
> I’d like to raise a process question regarding the application of the AI 
> policy to reviewed pull requests.
>
> A recent  PR  addressing a long-standing performance issue in degree() was 
> closed with the label “AI slop,”

I definitely should have been more polite so I will acknowledge that.
I guess I must have been getting annoyed with there being so many low
quality and AI PRs.

If you want to review and merge that PR then go ahead.

> The author disclosed AI usage and stated that finding the source of the 
> problem and design were manual
>
> The PR appears to comply with current AI policy as written (or at least I 
> don't see the violation)

I think that this is a bit disingenuous and that you know that the
code was all written by AI and that this was not honestly disclosed.
The AI policy says that you should explain how you have used AI and
that is in the PR template but what was written there was just "review
suggestions were provided by an AI tool". Maybe you read that
differently from me but what it should honestly say is "the code was
all written by Claude".

Much like I can see a few lines of code and know immediately that it
was written by Christopher Smith I can also see a few lines of code
and know that it was written by Claude or its ilk. I'm pretty sure
that you can also read a few lines of code and say the same things.

> As the AI policy currently stands, it permits AI-assisted contributions 
> provided the author understands and takes responsibility for the code. And 
> having reviewed the code, I can't see why labelling it as “AI slop” is a 
> sufficient basis for closure in the absence of technical objections.

I think that you are misunderstanding what I would consider to be the
problem of AI PRs. It is not about the actual code and its quality.
Many AI tools can produce better code than many of the people
currently opening SymPy PRs. The problem with the AI PRs is that they
are harder to review and are overloading us with spam. The problem
also is that if people don't explain honestly how they have used AI
then that in itself makes it harder to review the code because you
have to pick apart the AI hallucinations from the human
misunderstandings.

The other problem is that if we really want AI PRs then we don't
actually need new contributors to bring them. It would be far more
efficient for those of us who would have reviewed the PR to make AI
PRs directly without the other person getting in the way.

Technically the main problem with the PR is just the fact that it is
classic more code on top of code creating more space for bugs without
actually delivering much value. Unfortunately this is exactly what AI
makes easy: you ask it to do something and it just spits out more and
more code. The code might seem to work but if we merge it into the
codebase then it needs to be maintained and soon there will be bug
reports saying that degree doesn't work in this or that corner case
and then someone will have to review the bug reports and then someone
else will have an AI spit out more code on top of code and so on.

The PR may "fix" some issue but how much actual value are you getting?
It doesn't actually make it safe to call degree on an expression
because it still has a fallback case where it would effectively go
into an infinite loop like degree((x - 1)**1000000 - (x +
1)**1000000). The proper fix would be to have a version of Poly that
does not expand everything and that would be something very useful
that could be used in lots of places for much more than just degree
like it could also get the leading coefficient, support different
domains, evaluate and so on.

In any case all of the low quality PRs around now are seriously
getting me down. I think for now I will stop reviewing PRs from anyone
but a select list of people whose PRs don't annoy me (people who can
consistently produce something that does not require many iterations
unless it delivers significant actual value). This will mean that
unless other people do a lot more PR reviewing most GSOC-related PRs
are not going to get reviewed.

--
Oscar

-- 
You received this message because you are subscribed to the Google Groups 
"sympy" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion visit 
https://groups.google.com/d/msgid/sympy/CAHVvXxTi4x5ZBAeUmWhLZsU%2B_ffYFCL7Y87HnKhRAjMd0XL3sQ%40mail.gmail.com.

Reply via email to