[FFmpeg-devel] Fw: GSoC 2025: Please review your Project Ideas lists

[Soft Works]

> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of
> Michael Niedermayer
> Sent: Montag, 17. Februar 2025 01:25
> To: FFmpeg development discussions and patches <ffmpeg-
> devel at ffmpeg.org>
> Subject: Re: [FFmpeg-devel] Fw: GSoC 2025: Please review your Project Ideas
> lists
> 
> Hi Soft Works
> 
> On Sun, Feb 16, 2025 at 07:42:11PM +0000, Soft Works wrote:
> >
> >
> > > -----Original Message-----
> > > From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of
> > > Michael Niedermayer
> > > Sent: Freitag, 14. Februar 2025 20:29
> > > To: FFmpeg development discussions and patches <ffmpeg-
> > > devel at ffmpeg.org>
> > > Subject: [FFmpeg-devel] Fw: GSoC 2025: Please review your Project Ideas
> lists
> > >
> > > Hi everyone
> > >
> > > Please help improve the gsoc 2025 page
> > >
> > > 2 ideas are missing backup mentors, if you are able please
> > > add yourself to them as backup mentor
> > >
> > > Also please add more ideas!
> >
> > Hi Michael,
> >
> > would this be a valid gsoc project?
> >
> >
> > - Develop a scraping tool which iterates ffmpeg forks on GitHub, investigates
> >   branches and identifies patches which don't exist in upstream ffmpeg
> > - Deduplicate results
> > - Setup a customized AI with knowledge about the ffmpeg source code
> > - Run this AI to classify the results:
> >   - feature or fix
> >   - generally valid or just special cake
> >   - value for the project
> > - Finally filter and produce a list of high-value patches which are worth
> looking at
> >
> > I believe there are many good and valuable patches out there, which never
> made
> > It into ffmpeg because developers didn't want to go through the submission
> > Procedures.
> 
> My personal oppinion is that this is a great idea. And that it would be
> valuable for FFmpeg.
> If you checked teh GSoC rules and the community has no objections
> against it then it should be ok

TBH, I didn't check anything and I don't know much about GSoC. This is just an idea that I was carrying around for a while (waiting for the right moment to drop it 😊 )

> maybe reword the "customized AI" so that its clear that this could be
> something simple like just asking chatgpt to clasify commit messages.
> So it doesnt look like this would require building/training a new model
> from scratch. (as google complained about over speced tasks)

(What are speced tasks?)

I would make it that easy. What I had in mind when with "customized AI" was preparing the ffmpeg code file as "embeddings" (in a vector format). This makes it available to the model without being part of the input. 
How effective that would be  - I don't know, but I'm very curious about finding out.
Another option would be to at least perform a "casting" of models to find out which is most familiar with ffmpeg.

Should this work well, a follow-up project might be to build an automated pull request validation based on that.
Not that it would be required, things are always reviewed very carefully and critically, but it might find obvious things without anybody needing to spend time on it and it would give feedback to submitters early.

sw